Dynamical Systems Theory

Björn Birnir
Center for Complex and Nonlinear Dynamics
and Department of Mathematics
University of California
Santa Barbara1

1 c 2008, Björn Birnir. All rights reserved.

2
Contents

1 Introduction 9
1.1 The 3 Body Problem . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2 Nonlinear Dynamical Systems Theory . . . . . . . . . . . . . . . 11
1.3 The Nonlinear Pendulum . . . . . . . . . . . . . . . . . . . . . . 11
1.4 The Homoclinic Tangle . . . . . . . . . . . . . . . . . . . . . . . 18

2 Existence, Uniqueness and Invariance 25

2.1 The Picard Existence Theorem . . . . . . . . . . . . . . . . . . . 25
2.2 Global Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3 Lyapunov Stability . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.4 Absorbing Sets, Omega-Limit Sets and Attractors . . . . . . . . . 42

3 The Geometry of Flows 51

3.1 Vector Fields and Flows . . . . . . . . . . . . . . . . . . . . . . . 51
3.2 The Tangent Space . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.3 Flow Equivalance . . . . . . . . . . . . . . . . . . . . . . . . . . 60

4 Invariant Manifolds 65

5 Chaotic Dynamics 75
5.1 Maps and Diffeomorphisms . . . . . . . . . . . . . . . . . . . . . 75
5.2 Classification of Flows and Maps . . . . . . . . . . . . . . . . . . 81
5.3 Horseshoe Maps and Symbolic Dynamics . . . . . . . . . . . . . 84
5.4 The Smale-Birkhoff Homoclinic Theorem . . . . . . . . . . . . . 95
5.5 The Melnikov Method . . . . . . . . . . . . . . . . . . . . . . . 96
5.6 Transient Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . 99

6 Center Manifolds 103

3
4 CONTENTS

7 Bifurcation Theory 109

7.1 Codimension One Bifurcations . . . . . . . . . . . . . . . . . . . 110
7.1.1 The Saddle-Node Bifurcation . . . . . . . . . . . . . . . 111
7.1.2 A Transcritical Bifurcation . . . . . . . . . . . . . . . . . 113
7.1.3 A Pitchfork Bifurcation . . . . . . . . . . . . . . . . . . . 115
7.2 The Poincaré Map . . . . . . . . . . . . . . . . . . . . . . . . . . 118
7.3 The Period Doubling Bifurcation . . . . . . . . . . . . . . . . . . 119
7.4 The Hopf Bifurcation . . . . . . . . . . . . . . . . . . . . . . . . 121

8 The Period Doubling Cascade 123

8.1 The Quadradic Map . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.2 Scaling Behaviour . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.2.1 The Singularly Supported Strange Attractor . . . . . . . . 138

A The Homoclinic Orbits of the Pendulum 141

List of Figures

1.1 The rotation of the Sun and Jupiter in a plane around a common
center of mass and the motion of a comet perpendicular to the plane. 10
1.2 The nonlinear pendulum is your typical grandfather’s clock,
where the pendulum makes the angle θ with the vertical line. The
distance that the pendulum travels is the arc length `θ where ` is
the length of the arm and gravity pulls on the pendulum with the
force F1 = mg where m is the mass of the pendulum and g the
gravitational acceleration. . . . . . . . . . . . . . . . . . . . . . . 13
1.3 W u denotes the unstable and W s the stable manifolds, Eu is the
(linear) unstable subspace and Es the (linear) stable subspace. . . . 15
1.4 The phase portrait for the nonlinear pendulum shows four differ-
ent type of solutions. The first type are the stationary solutions
at the origin, that is a center and the one at (±π, 0) that is a sad-
dle. Around the origin there are periodic orbits corresponding to
small oscillations of the pendulum that are called librations. Then
there are two homoclinic connections connecting the saddle to it-
self, because (±π, 0) are really the same point corresponding to
the vertical position of the pendulum. The fourth type of solu-
tions are rotations, above and below the homoclinic connections.
They correspond to rotations of the pendulum around the circle
with increasing velocity as the distance from the origin increases. 17
1.5 The phase portrait of the damped pendulum is similar to the un-
damped one except that the origin has changed to a sink and all
solutions, except the saddle and the two rotations that connect to
the stable manifolds of the saddle, spiral into the sink. . . . . . . . 19

5
8 LIST OF FIGURES
Chapter 1

Introduction

1.1 The 3 Body Problem

In the 16th and 17th century major progress was made in celestial mechanics that
at the time constituted the theoretical part of astronomy. It started with the careful
observation made by Tycho Brahe which were then formulated into Kepler’s laws
by Kepler and thereafter lead to the solution of the two-body problem in astron-
omy. This was the integration of the equations of motions of two large celestial
bodies such as the Sun and Jupiter in our planetary system that can be formulated
as the rotation around a common center of gravity, see Figure 1.1. These develop-
ments lead to the great hope of the 18th century that the equations of motion for
our planetary system could be integrated and a great deal of effort was expended
trying to integrate the three-body problem. This was the model problem for the
motion of two bodies such as the Earth and Jupiter around the Sun, ignoring all
the other planets. In spite of a great effort of most of the best mathematicians of
the time no progress was made for most of the 18th century.
The methods that most of the mathematicians used to try to solve the problem
was to find additional integrals of the motion. Since the three-body problem is a
Hamiltonian system it was known that if enough integrals of the motion could be
found then the equations of the motion could be integrated.
In 1898 Poincaré dropped a bombshell that shook the mathematical world.
He proved that for a three-body problem consisting of the Sun and Jupiter and a
comet moving perpendicular to the planetary plane, see Figure 1.1, there exist no
analytic integrals. In proving this he set the stage for mathematics of the latter
half of the 19th century and discovered Nonlinear Dynamics!

9
10 CHAPTER 1. INTRODUCTION

Figure 1.1: The rotation of the Sun and Jupiter in a plane around a common center
of mass and the motion of a comet perpendicular to the plane.
1.2. NONLINEAR DYNAMICAL SYSTEMS THEORY 11

1.2 Nonlinear Dynamical Systems Theory

Nonlinear dynamics has profoundly changed how scientist view the world. It
had been assumed for a long time that determinism implied predictability or if
the behavior of a system was completely determined, for example by differential
equation, then the behavior of the solutions of that system could be predicted for-
ever after. Nonlinear dynamics showed that this assumption was false. Namely,
one could have a solution that was completely determined by an ordinary differ-
ential equation (ODE) and an initial condition and still its trajectory could not be
predicted. The reason is that the nonlinearity leads to an instability which in turn
implies a loss of predictability. This effect is best described by:

The Butterfly Effect: A monarch butterfly fluttering its wings in Santa Bar-
bara, California, today may cause a storm in the town of Akureyri, on the north
coast of Iceland, in two weeks time.

What happens is that the instability leads to a loss of information and this is
the perfect nightmare of someone who is trying to use a computer to obtain his
answers. The tiny round-off errors that one finds in any numerical computation
by a computer are being magnified exponentially in time. This means that after a
short time the error in the computation is overwhelming the answer and the result
of the computation is numerical junk.
We will in later chapters associated chaotic behavior with positive Lyapunov
exponents and another way of saying the above is that positive Lyapunov expo-
nents lead to limits on predictability.

1.3 The Nonlinear Pendulum

The motion of the nonlinear pendulum, see Figure 1.3 is determined by Newton’s
law
F = ma
where m is the mass and a the acceleration. Now the arclength that the pendulum
travels is `θ, where ` is the length of the arm of the pendulum and θ is the angle
from the vertical directions. From Figure 1.3 we can tell using basic trigome-
try that F = −mg sin(θ) where the gravitational force is F1 = mg where g is the
gravitational acceleration. The negative sign in F comes from the fact that θ is
measured in the counterclockwise direction, whereas F points in the clockwise
12 CHAPTER 1. INTRODUCTION

direction. The force F2 in Figure 1.3 is balanced by the tension in the arm of the
pendulum. Thus the above equation becomes
m`θ̈ = −mg sin(θ)
Dividing by m` we get that
g
θ̈ + sin(θ) = 0
`
where r
g
ω=
`
is the frequency of the nonlinear pendulum. The energy of the nonlinear pendulum
is obtained by multiplying the equation by θ̇
θ̇θ̈ + ω2 sin(θ)θ̇ = 0
and integrating with respect to t
θ̇2
+ ω2 (1 − cos(θ)) = C
2
where C is a constant. If we damp and drive the nonlinear pendulum its motion is
described by the equation
θ̈ + δθ̇ + ω2 sin(θ) = ε cos(Ωt)
Now we let x = θ and set ω = 1, then the nonlinear pendulum (without damp-
ing and driving) is described by the equation
(1.1) ẍ + sin (x) = 0
x(0) = x0 , ẋ(0) = ẋ0 .
x0 and ẋ0 are the initial position (angle) and initial (angular) velocity of the pendu-
lum. With these initial conditions specified, the initial value problem (IVP) (1.1)
determines the solution x(t) for all time. We will now give a complete qualitative
analysis of the solutions of the Equation 1.1. This is done by the following steps.
1. Write the equations as a first order system.
We let y = ẋ and rewrite the equation as a first order system
   
d x y
= .
dt y − sin (x)
1.3. THE NONLINEAR PENDULUM 13

Figure 1.2: The nonlinear pendulum is your typical grandfather’s clock, where the
pendulum makes the angle θ with the vertical line. The distance that the pendulum
travels is the arc length `θ where ` is the length of the arm and gravity pulls on
the pendulum with the force F1 = mg where m is the mass of the pendulum and g
the gravitational acceleration.
14 CHAPTER 1. INTRODUCTION

2. Find the stationary solutions.

If we set the right hand side of last equation equal to zero, we get two
equations
y = 0, sin(x) = 0,
that imply that the origin (x, y) = (0 (mod 2π), 0) is a stationary solution
and so is (x, y) = (π (mod2π), 0). The straight-down position of the pendu-
lum (x, y) = (0, 0) is clearly a stationary solution. It should be stable. The
straight-up position of the pendulum (x, y) = (±π, 0) is also a stationary so-
lution. It ought to be unstable. The phase plane has the period 2π in the x
direction so it is really a cylinder.

3. Determine the stability of the stationary solutions

Consider the nonlinear pendulum

ẍ + sin (x) = 0.
∂x
We linearize the system and let z = ∂x o
, the derivative of x with respect to
the initial condition xo . Then the linearized system becomes

ż = D(x,y) f (x)z

where f (x, y) is the vector field

 
y
f (x, y) =
− sin (x)

Now D(x,y) f is simply the Jacobian of f

   
y 0 1
D(x,y) =
− sin (x) − cos (x) 0

We first evaluate Jacobian at the stationary solution

   
0 0 1
D(x,y) f =
0 −1 0

The eigenvalues of this matrix are λ = ±i, with no real eigenvectors. This
says that the stationary solution (x, y) = (0, 0) is marginally stable. Next we
1.3. THE NONLINEAR PENDULUM 15

Figure 1.3: W u denotes the unstable and W s the stable manifolds, Eu is the (linear)
unstable subspace and Es the (linear) stable subspace.
16 CHAPTER 1. INTRODUCTION

evaluate the Jacobian at the stationary solution (x, y) = (π (mod 2π), 0) to

get the matrix  
0 1
.
1 0

√ eigenvalues of √this matrix are λ = ±1, with eigenvectors

The
(1/ 2)(1, +1) and (1/ 2)(1, −1) respectively. That one of the eigen-
values has a positive real part implies that the stationary solution (x, y) =
(π (mod 2π), 0) is unstable. The eigenvectors determine the linear unstable
and stable manifolds Eu and Es , see Figure 3, and Theorem 2.2 give the ex-
u (π, 0) and stable W s (π, 0) manifolds,
istence of the nonlinear unstable Wloc loc
which are tangent to Eu and Es respectively, see Figure 3.

4. Draw the phase portrait

The phase portrait is shown in Figure 4.

5. Describe the different types of solutions.

(a) The first type are the above stationary solutions.

(b) The circles and ellipses around the origin represent small oscillations
of the pendulum around the straight-down stationary solution. These
types of solutions are called librations.
(c) The homoclinic connections connecting the straight up (π, 0) station-
ary solution with itself are the third type of solutions.
(d) The rotation above the homoclinic connections represent solutions that
rotate around the circle in two possible directions.
(e) There are only these four types of solutions.

If we multiply the equation (1.1) by ẋ and integrate with respect to t we get

the energy of the pendulum

y2
E(x, y) = + 1 − cos (x) = constant.
2
This is obviously the same expression as we obtained above with θ = x, θ̇ = y and
ω = 1. The first part of this expression is the kinetic energy and the second part
the potential energy. E is constant for each solution (orbit) of the IVP, and it can
be used to prove the existence of the elliptical orbits around the origin in Figure
4. In this case, since the ODE is two-dimensional and has an integral, we can
1.3. THE NONLINEAR PENDULUM 17

Figure 1.4: The phase portrait for the nonlinear pendulum shows four different
type of solutions. The first type are the stationary solutions at the origin, that is a
center and the one at (±π, 0) that is a saddle. Around the origin there are periodic
orbits corresponding to small oscillations of the pendulum that are called libra-
tions. Then there are two homoclinic connections connecting the saddle to itself,
because (±π, 0) are really the same point corresponding to the vertical position
of the pendulum. The fourth type of solutions are rotations, above and below the
homoclinic connections. They correspond to rotations of the pendulum around
the circle with increasing velocity as the distance from the origin increases.
18 CHAPTER 1. INTRODUCTION

find W u and W s explicitly. The homoclinic orbits approach the stationary solution
(x, y) = (π (mod 2π), 0) as t → ∞ so by continuity they must have the same energy
as this stationary solution. We therefore set the energy equal to the energy of the
stationary solution
y2
+ 1 − cos (x) = E(π, 0) = 2.
2
This gives a first order equation for x that is solved, see Appendix A, to give

(x, y) = ±2(tan−1 [ sinh(t)], sech(t)),

the − sign gives W u (π, 0) and the + sign gives W s (π, 0). Notice that both of these
manifolds are globally defined.
Adding damping the pendulum equation (1.1) gives the initial value problem:

(1.2) ẍ + δẋ + sin (x) = 0

(1.3) x(0) = x0 , ẋ(0) = ẋ0 .

The phase portrait of its solutions is shown in Figure 1.3. The stability of the
stationary solution (x, y) = (π (mod 2π), 0) remains the same, this is because it
is hyperbolic and therefore structurally stable. The stationary solution (x, y) =
(0, 0) on the other hand turns into a sink. The homoclinic connections are not
structurally stable and break under the perturbation but one rotation from each
direction connects with the stable manifolds of the stationary solution (x, y) =
(π (mod 2π), 0). Every other solution eventually spirals into the sink, see Figure
1.3.

1.4 The Homoclinic Tangle

The analysis of the damped and driven nonlinear pendulum

(1.4) ẍ + δẋ + sin (x) = ε cos(ωt)

(1.5) x(0) = x0 , ẋ(0) = ẋ0 .

is more involved. The first problem is that the equation is non-autonomous and
one must analyze the extended phase space, which is three-dimensional, instead
of the two-dimensional phase space above. This is harder but more interesting
things can happen in three dimensions. Poincaré got around this problem by in-
venting the Poincaré map, see Figure 1.4. This allowed him to reduce the analysis
1.4. THE HOMOCLINIC TANGLE 19

Figure 1.5: The phase portrait of the damped pendulum is similar to the undamped
one except that the origin has changed to a sink and all solutions, except the saddle
and the two rotations that connect to the stable manifolds of the saddle, spiral into
the sink.
20 CHAPTER 1. INTRODUCTION

Figure 1.6: The plane perpendicular to the periodic orbit is called a transversal.
The Poincaré map is the return map to the transversal. Thus the Poincaré map
of the point q is the point p. The periodic orbit itself gives a fixed point of the
Poincaré map.
1.4. THE HOMOCLINIC TANGLE 21

of a three-dimensional vector field to the analysis of the two-dimensional map.

It will be our main tool in the analysis of chaotic dynamics. When ε is turned
on in the equation (1.4) the sink at the origin turns into a stable periodic orbit.
The hyperbolic point at (π(mod 2π), 0) turns into a hyperbolic (unstable) periodic
orbit. The orbits (they are now a sequence of points) in at neighborhood of the
origin still spiral into it, but something strange can happen in a neighborhood of
the hyperbolic fixed point, corresponding to the hyperbolic periodic orbit. The
hyperbolic fixed point still has stable W s and unstable W u manifolds. But if the
balance of the damping and the driving (ε and δ) is right the unstable manifold
can come into a neighborhood of the hyperbolic fixed point and cross its stable
manifold transversely,
W u (π, 0) ⊥ W s (π, 0)
see Figure 1.4. This could not happen in a flow because it would violate the
uniqueness of the solution starting at the intersection point. But it can happen in a
map because then the intersection point just corresponds to a particular orbit of the
flow. If these manifolds cross once they must cross infinitely often and we get a
homoclinic tangle close to the hyperbolic fixed point, see Figure 1.4. This is what
Poincaré discovered in the three-body problem and he showed that the existence
of the homoclinic orbit implies that no analytic integrals of the motion can exist.
We will go through his argument when we study the Smale horseshoe in Chapter
5.

Exercise 1.1
Perform the qualitative analysis of the Duffing’s equation,

ẍ − x + x3 = 0,

1. Find the stationary solutions.

2. Determine the stability of the stationary solutions.

3. Draw the phase portrait of the Duffings equation.

4. Use the phase portrait and (1) - (3) to identify four different types of solu-
tions of the Duffing’s equation and describe their qualitative behaviour.

5. Find the energy of the Duffing’s equation and use it to compute the homo-
clinic orbits, compare Appendix A.
22 CHAPTER 1. INTRODUCTION

Figure 1.7: The phase space of the Poincaré map of the damped and driven non-
linear pendulum.
1.4. THE HOMOCLINIC TANGLE 23

6. Add damping to the equation

ẍ = δx − x + x3 = 0,

and describe how this changes the analysis (1) - (3) above when δ is small.

7. Add forcing to the equation

ẍ = δx − x + x3 = ε cos(ωt),

and describe what it’s Poincaré map looks like for small ε.
24 CHAPTER 1. INTRODUCTION
Chapter 2

Existence, Uniqueness and

Invariance

2.1 The Picard Existence Theorem

The fundamental theorem in ordinary (and partial) differential equations is the
Picard existence theorem. We start with the definition of Lipschitz continuous
functions.

Definition 2.1 A function f : U × [t0 ,t0 + T ] → Rn , U ⊂ Rn , is Lipschitz continu-

ous in U if there exists a constant L such that

k f (y,t) − f (x,t)k ≤ Lky − xk

for all x, y ∈ U and t ∈ [t0 ,t0 + T ]. If U = Rn , f is called globally Lipschitz.

Theorem 2.1 Consider the initial value problem

ẋ = f (x,t), x(t0 ) = x0 , x ∈ Rn , t ∈ R, (2.1)

in the time interval |t − t0 | ≤ a, and the box x ∈ D, D = {x|kx − x0 k ≤ d}, here

a and d are positive constants. Suppose that the function (vector field) f (x,t)
satisfies the two conditions

(i) f (x,t) is continous in G = I × D = [t0 − a, t0 + a] × D,

(ii) f (x,t) is Lipschitz continous in G,

25
26 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

then there exists a unique solution of (2.1) for the time interval |t − t0 | ≤
d
max (a, M ), where
M = max k f (x,t)k.
G

Proof: We write the ODE and the initial condition (2.1) as the integral equation
Z t
x(t) = x0 + f (x(s), s)ds. (2.2)
t0

It is easy to see that (2.1) and (2.2) are actually equivalent. Picard’s idea was to
iterate this equation and define the sequence
Z t
x1 (t) = x0 + f (x0 , s)ds
t0
Z t
x2 (t) = x0 + f (x1 (s), s)ds
t0
..
. Z t
xn+1 (t) = x0 + f (xn (s), s)ds
t0
..
.

If the sequence of iterates {xn (t)} converges it will converge to the solution of
(2.2). The first thing we must check is that the iterates actually lie in D, see Figure
2.1,
Z t
kx1 (t) − x0 k = k f (x0 , s)dsk
t0
Z t
≤ k f (x0 , s)kds
t0
≤ M|t − t0 | ≤ d
d
for |t − t0 | ≤ min (a, M ). This was the reason for the choice of the length of the
time interval, as this minimum. Now the computation is the same for all the
iterates so the whole sequence lies in D.
Next we use induction to show that successive iterates satisfy the inequality

MLn−1 |t − t0 |n
kxn (t) − xn−1 (t)k ≤ .
n!
2.1. THE PICARD EXISTENCE THEOREM 27

Figure 2.1: The square |t − to | ≤ a, kx − xo k ≤ b, where the solution of the initial

value problem (2.1) exists.
28 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

When n = 1 this is the estimate above

kx1 (t) − x0 k ≤ M|t − t0 |.

We assume
MLn−2 |t − t0 |n−1
kxn−1 (t) − xn−2 (t)k ≤
(n − 1)!
and consider the difference of
Z t
xn (t) = x0 + f (xn−1 (s), s)ds
t0

and Z t
xn−1 (t) = x0 + f (xn−2 (s), s)ds,
t0
Z t
xn (t) − xn−1 (t) = [ f (xn−1 (s), s) − f (xn−2 (s), s)] ds
t0
so
Z t
kxn (t) − xn−1 (t)k ≤ k f (xn−1 (s), s) − f (xn−2 (s), s)kds
t0
Z t
≤ L kxn−1 (s) − xn−2 (s)kds
t0

by the Lipschitz condition

|s − t0 |n−2 |s − t0 |n
Z t
n−1
≤ ML ds = MLn−1 t
t0
t0 (n − 1)! n!
|t − t0 |n
= MLn−1
n!
This completes the induction.
Now the space of continous funtions on the interval I is a complete metric
space with the metric
kx(t)ksup = max |x(t)|
I
and to show that the sequence {x(t)} converges we just have to show that it is
Cauchy, i.e. given ε > 0, ∃N such that

kxn (t) − xm (t)k < ε, for ∀ n, m ≥ N.

2.1. THE PICARD EXISTENCE THEOREM 29

But this is straight-forward

n
kxn (t) − xm (t)k ≤ ∑ kx j (t) − x j−1 (t)k
j=m+1
n
L j−1 |t − t0 | j
≤ M ∑
j=m+1 j!

by the estimate above

M ∞
L j |t − t0 | j M  L|t−t0 | 
≤ ∑ j! = e − 1 .
L j=1 L

If we choose |t − t0 | sufficiently small then

M  L|t−t0 | 
kxn (t) − xm (t)k ≤ e − 1 < ε,
L
for ∀ n, m. This means that {xn (t)} is a Cauchy sequence and converges to a con-
tinous function on I, xn (t) −→ x(t) as n −→ ∞. Moreover, taking the limit as
n −→ ∞ in Z t
xn (t) = x0 + f (xn−1 (s), s)ds,
t0
we get Z t
x(t) = x0 + f (x(s), s)ds
t0
by uniform convergence, so x(t) satisfies (2.2). This in turn implies that x(t) is
continously differentiable since f is continous so x(t) satisfies (2.1). QED

Example 2.1 Consider the IVP

ẋ = x1/2
x(0) = 0.

x ≡ 0 is a solution but so is
(t − c)2 /4, t ≥ c > 0

x(t) =
0, t < c.

We have infinitely many solutions one for each value of c, see Figure 2.1.
30 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Figure 2.2: The top figure show the graphs of two of the many solutions that
satisfy the ODE in Example 2.1. The bottom figure show the solution in Example
2.2, that is blowing up in a finite time.
2.1. THE PICARD EXISTENCE THEOREM 31

Example 2.2 Finite-time Blow-up

Consider the IVP

ẋ = x2
x(0) = x0 .

The solution is
x0
x(t) = .
(1 − x0t)
It exists for t < 1/x0 , but
lim x(t) = +∞,
t→1/x0

see Figure 2.1.

Exercise 2.1

1. Show that the IVP

p
ẋ = 1 − x2
x(0) = x0

has infinitely many continuous solutions in the region |x| ≤ 1. Show also
that there are only two solutions that have a continuous second derivative ẍ.

2. Find the solution to the IVP

ẋ = 1 + x2
x(0) = x0

and the finite time when it blows up.

The following (more general) form of Grönwall’s inequality is an important

technical tool.

Theorem 2.2 Let u(t) and v(t) be non-negative continuous functions on [α, β],
C ≥ 0 a constant, and suppose
Z t
v(t) ≤ C + v(s)u(s)ds, for α ≤ t ≤ β. (2.3)
α
32 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Then
t
Z 
v(t) ≤ C exp u(s)ds , for α ≤ t ≤ β (2.4)
α

and if C = 0, v(t) ≡ 0.

Proof: We let
Z t
C+ v(s)u(s)ds = V (t).
α
Then
v(t) ≤ V (t)
by the hypothesis and we assume in addition that

v(t) > 0, for (t − α) small,

then
V 0 (t) = v(t)u(t) ≤ V (t)u(t).
We can solve this differential inequality, because

v(t) > 0, for (t − α) small ⇒ V (t) > 0.

The solution is
t t
Z  Z 
V (t) ≤ V (0) exp u(s)ds = C exp u(s)ds .
α α

This implies that

t
Z 
v(t) ≤ C exp u(s)ds
α

and
C = 0 ⇒ v(t) ≡ 0 for (t − α) small.
But then v(t) ≡ 0, for all t ∈ [α, β] because we can make the same argument with
any α1 ∈ [α, β] instead of α. QED

The first application shows that a solution to an ODE depends continuously

on its initial data, for finite time.
2.1. THE PICARD EXISTENCE THEOREM 33

Corollary 2.1 Let y(t) and x(t) be two solutions of the ODE (2.1), with a Lips-
chitz continuous vector field f and assume that the solution exists for |t −t0 | < T0 .
Then for ε > 0 and T < T0 , there exists a δ > 0 such that

ky0 − x0 k < δ ⇒ ky(t) − x(t)k < ε

for |t − t0 | ≤ T .

Proof: We use the integral equation (2.2),

Z t
y(t) − x(t) = y0 − x0 + ( f (y(s), s) − f (x(s), s))ds
t0

and the Lipschitz continuity to get the inequality

Z t
ky(t) − x(t)k ≤ ky0 − x0 k + K ky − xk(s)d|s|.
t0

Then Theorem 2.2 states that

ky(t) − x(t)k ≤ ky0 − x0 keK|t−t0 | .

Now
ky0 − x0 k < εe−Kt = δ
implies that
ky(t) − x(t)k < ε, for |t − t0 | < T.
QED

The next application of the Grönwall’s inequality is to show that if the deriva-
tive of the vector field is globally bounded, then the solution is bounded by an
exponential function. First we must discuss:

Theorem 2.3 The Mean-Value Theorem in Rn

If f (x,t) is C1 in Rn , that is the Jacobian derivative Dx f exists and is contin-

uous, then
Z 1
f (y,t) − f (x,t) = Dx f (x + s(y − x),t) · (y − x)ds. (2.5)
0
34 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Proof: We apply the fundamental theorem of calculus

f (y,t) − f (x,t) = f (x + s(y − x),t)|s=1
s=0
Z 1
d
= f (x + s(y − x),t)ds
0 ds
Z 1
= Dx f (x + s(y − x),t) · (y − x)ds
0

where s is a parameter and · denotes the multiplication of an n × n matrix by a

vector in Rn . QED
Corollary 2.2 Suppose the derivative Dx f of the vector field in the IVP (2.1)
and (2.2) is bounded in a strip U = {(x,t) | x ∈ Rn , |t − t0 | ≤ d}. Let N =
maxU kDx f (x,t)k and M = sup|t−t0 |≤d k f (x0 ,t)k. Then the solution is bounded
M N|t−t0 |
kx(t) − x0 k ≤ (e − 1),
N
by an exponential function in U.
Proof: We use the integral equation (2.2)
Z t
x − x0 = f (x(s), s)ds
t0
Z t Z 1 
= f (x0 , s) + Dx f (x + r(x − x0 ), s) · (x − x0 )dr ds,
t0 0

by the mean-value theorem. Therefore

Z t Z 1 
kx − x0 k ≤ kDx f (x + r(x − x0 ))kdr · kx − x0 k + k f (x0 , s)k ds
t0 0
Z t 
M
≤ N kx − x0 k(s) + ds
t0 N
Thus Z t 
M M M
kx − x0 k + ≤ + N kx − x0 k(s) + ds
N N t0 N
and by Theorem 2.2
M M
kx − x0 k + ≤ exp (N|t − t0 |).
N N
This proves the assertion. QED
2.2. GLOBAL SOLUTIONS 35

If the vector field is only assumed to be continuous the solution of the Initial
Value Problem (2.1), but it may not be unique. The proof requires the classical
Arzela-Ascoli Theorem: A uniformly bounded and equicontinuous sequence of
functions on a compact set G ⊂ Rn+1 has a uniformly convergent subsequence.

Theorem 2.4 The Peano Existence Theorem

Let U ⊂ Rn × R be an open set and f (x,t) a continuous function on U. Then
there exists a constant c > 0 and a continuous function x(t), such that for every
(x0 ,t0 ) ∈ U, (x(t),t) ∈ U for |t − t0 | < c and x(t) satisfies the IVP

ẋ = f (x,t), x(t0 ) = x.

2.2 Global Solutions

We saw in Corollary 2.2 that if the derivative Dx f (x,t) of the vector field f is
bounded for all x and t then the solution of the IVP has exists globally. It is very
restrictive to require Dx f to be globally bounded and a less restrictive method is
to seek an a priori bound on x.

Example 2.3
The motion of a spring with a stiffening (nonlinear) restoring force is described
by the equation
ẍ = −x − x3
We move the two terms on the right hand side left and multiply by ẋ. Then inte-
gration in t yields the energy

ẋ2 x2 x4
+ + = constant = K
2 2 4
This implies that
|x| ≤ 2K 1/2 , and |ẋ| ≤ (2K)1/2
for all time, so the solution exist globally, see Corollary 2.3 below.

Exercise 2.2 Show that the solutions of the nonlinear pendulum

ẍ + sin(x) = 0

have global existence.

36 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Exercise 2.3 What can you say about the interval of existence of the solution to

ẍ + x2 = c

where c is a constant?
Global solutions are obtained by piecing together local solutions. In fact it
follows from the Local Existence Theorem 2.1 that
1. Every point (to , xo ) lies on some solution (t, x(t)).

2. If (to , xo ) lies on two solutions (t, x(t)) and (t, y(t)), then these solutions
must be the same in an interval |t − to | < a, around to .
The second statement above is strengthen considerably by the follwing Lemma

Lemma 2.1 If two solutions of the same ODE agree at a point then they must be
the same in their whole interval of existence.

Proof: Suppose that the solution (t, x(t)) is defined on the interval [a, b) and the
other solution (t, y(t)) is defined on the interval [a, c) with b < c. Let

d = in f {t|x(t) 6= y(t)}

and consider the interval [a, d]. x(t) and y(t) both exist on [a, d) and since they are
continuous, their limits at d must be equal

lim x(t) = lim y(t)

t→d t→d

This means that d > b and consequently b = c because otherwise x(t) exists be-
yond b contradicting the statement above. QED

We define the solution in the large (t, z(t)) to be the union of the local solutions
(t, x(t)). It exists on the union of the local intervals and is uniquely defined by any
initial point (to , xo ) by above discussion.
Next we prove in two steps that if solution only exist for a finite time interval
then they must blow-up.

Theorem 2.5 Let D̄ be a closed and bounded subset of a region U where solution
(t, x(t)) is defined. If the solution is only defined on the time interval [a, b) with
b < ∞, then (t, x(t)) must leave D̄ for all t sufficiently close to b.
2.2. GLOBAL SOLUTIONS 37

Figure 2.3: The compact set D̄ has the distance γ to the compliment of the region
U.
38 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Proof: We assume there exists a solution (t, x(t)) defined for 0 ≤ |t − t0 | < β and
passing through some point (t0 , x0 ) of D̄. We have to prove that there exists an
α such that the solution lies outside D̄, for β − α < t − t0 < β. Now consider the
region Rn+1 \U, see Figure 2.2. Let

γ = inf{(x,t)∈D̄,(y,τ)∈Rn+1 \U} (|t − τ| + kx − yk)

be the distance between D̄ and Rn+1 \U, and consider the rectangles

R(τ, y) = {(t, x)||t − τ| ≤ δ, kx − yk ≤ δ}

where δ < γ. Then

D∗ = ∪(τ,y)∈D̄ R(τ, y)

is also a closed and bounded set. Now let

α = min(δ, δ/M), M = max(t,x)∈D∗ k f k.

Then by the Picard Existence Theorem 2.1 (t, x(t)) exists in the interval |t − t0 | <
α. Now if β − α < t1 < β and (t1 , x(t1 )) ∈ D∗ then the solution would exist beyond
t − t0 = β which is a contradiction. QED

Corollary 2.3 If the solution x(t) of the IVP only exists for time less than b < ∞,
then
lim kx(t)k = ∞
t→b

Proof: We let U in Theorem 2.5 be the strip

U = {(t, x)||t| < b}

and D be the rectangle

D = {(t, x)||t| < b, kxk ≤ d}

Then (t, x(t)) must leave D̄ for t approaching b, but since it cannot leave the side
t = b of D, it must leave the sides kxk = d. However, d is arbitrary. QED
2.3. LYAPUNOV STABILITY 39

2.3 Lyapunov Stability

A first order system
ẋ = f (x) (2.6)
x(t0 ) = x0
is said to be autonomous if the vector field f (x) is independent of time. If ẋ = 0,
x is called a stationary solution of (2.6). Then x must be a singular point f (x) = 0
of the vector field. A solution is said to be Lyapunov stable if solutions that start
close to it stay close for all time.

Definition 2.2 A solution of the first order system (2.6) is Lyapunov stable if for
ε > 0, there exists δ > 0, such that

ky0 − x0 k < δ ⇒ ky(t) − x(t)k < ε,

for t ≥ to . x(t) is asymptotically Lyapunov stable if

lim y(t) = x(t).

t→∞

The tool that allows us to prove Lyapunov stability is a Lyapunov function.

Definition 2.3 A function V : Rn → R is called a Lyapunov function if it has con-

tinuous partial derivatives

1. V is non-negative, V (x) ≥ 0

2. V (x) = 0 if and only if x = x̄, a stationary solution

3. V is non-increasing
dV
≤0
dt
Theorem 2.6 Suppose the vector field f (x) is continuous in a neighborhood of
a stationary solution x̄ and that there exists a Lyapunov function for (2.6) in this
neighborhood. Then x̄ is stable. Morover, if
dV
<0
dt
then x̄ is asymptotically stable.
40 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

We define exit points and prove a technical lemma before proving the theorem.

Definition 2.4 Let f (x,t) be continuous on an open set U and D ⊂ U be an open

subset. (t, x0 ) ∈ ∂D ∩U is an exit point of D if (t, x(t)) ∈ D for t − ε < t < t0 and ε
small. If, moreover, (t, x(t)) 6∈ D̄, for t0 < t < t0 + ε, then (t0 , x0 ) is called a strict
exit point of D.

Lemma 2.2 Let f be continuous on an open subset U and D ⊂ U open such that
∂D ∩ U is either empty or consists of points which are not exit points. Then the
solution of (2.6), with (t0 , x0 ) ∈ D, stays in D as long as it exists.

Proof: Suppose that the solution (t, x(t)) leaves D at some time t1 , then

(t1 , x(t1 )) ∈ ∂D ∩U

and (t1 , x(t1 )) is an exit point contrary to hypothesis. QED

We now proof Theorem 2.6.

Proof: Choose ε > 0 and let γ < minkx−x0 k=ε V (x). We can then choose δ such
that the ball kx − x0 k ≤ δ lies inside the region V (x) = γ, see Figure 2.3. We define
the sets

U = {x| kx − x̄k < ε}, D = {x| V (x) < γ}, ∂D = {x| V (x) = γ}

and consider the function v(x) = V (x) − γ. Then

v̇ = V̇ ≤ 0

since V is a Lyapunov function. Therefore no point on ∂D is an exit point. Conse-

quently by Lemma 2.2, if (t0 , x0 ) ∈ D, then (t, x(t)) must stay inside D as long as
the solution exists. However, D ⊂ U so that kx(t) − x̄k < ε as long as the solution
exists, but this implies the x(t) must exist for all t, by Theorem 2.5, because x(t)
is prevented from running to the boundary of U.
Now suppose V̇ < 0, then V (x(t)) → 0 as t → ∞ and this implies that x(t) → x̄.
QED

Example 2.4
2.3. LYAPUNOV STABILITY 41

Figure 2.4: The set defined by V (x) = γ lies outside the δ ball and inside the ε ball.
42 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

The Lorenz equation

ẋ = σ(y − x)
ẏ = ρx − y − xz
ż = xy − βz

where σ, ρ and β are positive constants, describe Rayleigh-Bénard convection in

the atmosphere, see Lorenz [15]. This heating of air by the earth’s surface during
the day and cooling by the upper layers in the atmosphere is the basis mechanism
in the daily weather cycle. The x and y variables are two components of the
velocity of air and z is temperature. We show that the function

1
V (x, y, z) = (x2 + σ(y2 + z2 ))
2
is a Lyapunov function for the Lorenz equations. First V ≥ 0 is positive definite
since σ > 0, V (x, y, z) = 0 implies (x, y, z) = (0, 0, 0) and the origin is a stationary
solution of the Lorenz equations. This verifies the first two conditions of Defini-
tion 2.2. By use of the Lorenz equations

dV
= xẋ + σ(yẏ + zż)
dt
= σ(xy − x2 ) + σ(ρxy − y2 − xyz) + σ(−βz2 + xyz)

Thus

V̇ = −σ(x2 − (1 + ρ)xy + y2 ) − σβz2

σ(1 + ρ) 2 σ(1 − ρ) 2
= − (x + y2 ) − (x − y2 ) − σβz2
2 2
We conclude that V is a Lyapunov function if and only if ρ ≤ 0. This implies that
the origin is stable if ρ ≤ 0 and asymptotically stable if ρ < 0.

2.4 Absorbing Sets, Omega-Limit Sets and Attrac-

tors
If an ODE is dissipative then all orbits of solutions will end up in a bounded set
and one can make strong statements about the phase portrait.
2.4. ABSORBING SETS, OMEGA-LIMIT SETS AND ATTRACTORS 43

Definition 2.5 The ODE, ẋ = f (x), has an absorbing set D; if for every bounded
set U ⊂ Rn , there exists a time T (U) such that x(0) = xo ∈ U and t ≥ T (U) implies
that x(t) ∈ D.
This means that all orbits are eventually absorbed by D see Figure 2.4.

Example 2.5
Consider and ODE
ẋ + δx = f (x,t),
where k f (x,t)k ≤ K. We can integrate this equation to get

d δt
e x = eδt f (x,t)
dt
or Z t
−δt
x(t) = xo e + e−δ(t−s) f (x(s), s)ds
0
and
K
kx(t)k ≤ kxo ke−δt + (1 − e−δt ). (2.7)
δ
The last inequality implies that
K
kx(t)k ≤ +κ
δ
where κ is an arbitrarily small number, defines an absorbing set for the ODE.
Namely, if we solve the inequality (2.7) for t, we get
K K
kxo ke−δt + (1 − e−δt ) ≤ + κ
δ δ
and
K −δt
(kxo k − )e ≤ κ
δ
so
κ
−δt ≤ log[ ]
(kxo k − Kδ )
and
1 κ
t ≥ − log[ ].
δ (kxo k − Kδ )
44 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Figure 2.5: Every orbit starting in U is eventually absorbed by the absorbing set
D.
2.4. ABSORBING SETS, OMEGA-LIMIT SETS AND ATTRACTORS 45

This means that if the initial data lies in a bounded set defined by kxo k ≤ M, then
x(t) ∈ D for
1 κ
t ≥ T (M) = − log[ ].
δ (kxo k − Kδ )
Stationary solutions are the simplest structure that one can encounter in phase
space the ω and α limit sets are more general structures. Stationary solutions
are clearly invariant under the flow and the union of orbits tending towards them
in positive time is called their basin of attraction. We will now generalize these
notions to more complicated attracting sets.

Definition 2.6 The ω limit set of a solution to an ODE, x(t), x(0) = xo consists
of all points y such that there exists a sequence tn → ∞ as n → ∞ and

lim x(tn ) = y.
n→∞

The α limit set of x(t) consists of all points y such that there exists a sequence
tn → −∞ as n → ∞ and
lim x(tn ) = y.
n→∞

The ω limit set consists of all points that points sampled from the solutions
converge to in positive time, whereas the α limit set consists of the points that
points sampled from the solution converge to in negative time.

Theorem 2.7 The α and ω sets are closed sets, if the solution x(t) is also compact
then these sets are compact, non-empty and connected sets.

Proof: We first show that the ω limit set is closed. Suppose that

lim yn = y
n→∞

where yn ∈ ω{x(t)}. Then by definition of ω there exist sequences {tmn } such that

lim x(tmn ) → yn .
n→∞

We can choose a (diagonal) sequence {tnn } from the sequences {tmn } such that

lim x(tnn ) → y.
n→∞

Thus y ∈ ω{x(t)}.
46 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Next we prove the (sequential) compactness of ω{x(t)}. Let {yn } be a se-

quence in ω{x(t)}. Then there exist sequences {tkn }, such that {x(tkn )} converge
to yn as k → ∞. Since the orbit x(t) is compact the diagonal sequence {x(tnn )} has
j
a convergent subsequence {x(t j )}, that converge to y ∈ ω{x(t)} as j → ∞. This
j
implies that the corresponding subsequence y j = limk→∞ x(tk ) also converges to y
as j → ∞. We have shown that {yn } has a convergent subsequence.
Now ω{x(t)} is not empty if it is compact because we can find a convergent
subsequence of x(tn ) converging to a point
lim x(tk ) = y
k→∞

and then y ∈ ω{x(t)}. Thus ω{x(t)} contains at least one point y.

We now prove that ω{x(t)} is a connected set. Suppose that ω consists of
two disjoint sets X and Y . Then both X and Y must be compact and the distance
between them is a positive number
d(X,Y ) = d.
However since both X and Y consist of limit points of x(t), x(t) visits both X
and Y infinitely often for arbitrarily large t. Moreover, since x(t) is a continuous
function of t and x(t) is covering the distance between X and Y infinitely often,
there exists a sequence {x(tn ) = xn } such that
d
d(Y, xn ) = .
2
The sequence {x(tn )} has a convergent subsequence
lim xk = z
k→∞

and thus
lim x(tk ) = z,
k→∞
so z ∈ ω{x(t)}. The distance
d
d(X, z) ≥ d(X,Y ) − d(Y, z) = ,
2
so z is neither in X nor Y , but this contradicts the decomposition
ω = X ∪Y.
The proof for the α limit set is similar. QED
2.4. ABSORBING SETS, OMEGA-LIMIT SETS AND ATTRACTORS 47

Definition 2.7 The time-τ map of an ODE ẋ = f (x), x(0) = xo , is defined to be

Tτ x(t) = x(t + τ).
This is also called the time-advance map.

Definition 2.8 A set D is invariant under a map T if T D ⊂ D, it is negatively

invariant if D ⊂ T D. In particular, D is invariant if it is both positively and nega-
tively invariant.

Exercise 2.4 Show that ω(xo ) and α(xo ) are both invariant.

Definition 2.9 If D and A are both subsets of Rn and kT n D − Ak → 0 as n →

∞, then we say that A attracts D. If A attracts a neighborhood of itself and is
invariant, it is called an attractor; if it is also compact and attracts Rn , it is called
a global attractor.

Theorem 2.8 If an ODE, ẋ = f (x), has an absorbing set D, then the ω-limit set
of D
ω(D) = ∩n ∪m≥n T m D
is a global attractor.
Proof: We first observe that ω(D) is compact and non-empty because it is a
nested intersection of closed and bounded sets. Then we show that it is invariant.
Let Kn = ∪m≥n T m D and K = ω(D). Then K is positively invariant because if
y ∈ K, there exists a sequence zn = T n yn → y and T m zn = T m+n yn ∈ K m+n , thus
T m y = limn→∞ T m+n yn ∈ K.
K is also negatively invariant. If T n yn → y ∈ K then {T n−m ym } has a convergent
subsequence, by the compactness of K, and if z = limn→∞ T n−m ym , then
y = T m z ∈ T m K.
This shows that K ⊂ T m K. Now K attracts a neighborhood, because if it does not,
then there exists a sequence yn in a neighborhood of K such that
kT n yn − Kk ≥ ε > 0.
But {T n yn } has a convergent subsequence
y = limm→∞ T m ym ∈ K,
which is a contradiction. Moreover, K attracts D which attracts Rn so K attracts
Rn . QED
48 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

Exercise 2.5 Show that the global attractor is unique.

Theorem 2.9 If an ODE, ẋ = f (x), has a Lyapunov function; then its global at-
tractor is connected and can be expressed as

A = W u (S),

the unstable set of the stationary solutions x̄ ∈ S,

W u (S) = {x ∈ Rn : x(−t), x(0) = x, is defined for t ≥ 0 and x(−t) → S, as t → ∞}.

Moreover, if each stationary solution is isolated, then S if finite and

A = ∪N u
x̄ j ∈SW (x̄ j ).

We first proof the following Lemma,

Lemma 2.3 If the ODE, ẋ = f (x), x(0) = xo , has a Lyapunov function and xo ∈
Rn , then ω(xo ) ∈ S and if x(t), is a compact orbit then α(xo ) ∈ S, also.

Proof: Since the Lyapunov function V (x(t)) is non-increasing, V (x(t)) → c, a

constant, as t → ∞. Now v = V − c is also a Lyapunov function and this means
that if x(tn ) → y ∈ ω(xo ), as tn → ∞, y = x̄ ∈ S is a stationary solution. Now let
tn → −∞ such that tn−1 − tn ≥ 1, T be the time-τ map, and consider

V (x(tn−1 + t)) ≤ V (x(tn + t))

where the inequality follow from the properties of V . If

x(tn + t) → y,

then by the compactness of x(t)

V (x(tn + t)) → c.

Then
v =V −c
is also a Lyapunov function and v(y) = 0. This means that y = x̄ ∈ S. QED

We can now proof Theorem 2.9.

2.4. ABSORBING SETS, OMEGA-LIMIT SETS AND ATTRACTORS 49

Proof: By Lemma 2.3 α(x) ⊂ S, for each x ∈ A. Thus A ⊂ W u (S). Clearly, S ⊂ A

and since A attracts a neighboorhood of itself W u (S) ⊂ A also.
If all stationary solutions x̄ ∈ S are isolated then S is a finite set, by the com-
pactness of A, and
A = ∪N u
x̄ j ∈SW (x̄ j ).

QED

The Picard Theorem 2.1 is an example (the first in history) of a fixed point or
contraction mapping theorem.

Theorem 2.10 The Contraction Mapping Principle A contraction mapping

d(T x, Ty) ≤ θd(x, y), 0 < θ < 1,

on a complete metric space has a unique fixed point

T x = x.

Proof: We first prove the uniqueness. Suppose x and y are both fixed points
T x = x, Ty = y. Then

d(x, y) = d(T x, Ty) ≤ θd(x, y)

since 0 < θ < 1,

(1 − θ)d(x, y) ≤ 0 =⇒ d(x, y) = 0
so x = y. Next we prove the estimate

θn
d(T m x, T n x) ≤ d(T x, x), m > n
(1 − θ)

for the sequence of (Picard) iterates {T n x}.

d(T m x, T n x) ≤ d(T m x, T m−1 x) + d(T m−1 x, T m−2 x)

+ · · · + d(T n+1 x, T n x),

by the triangle inequality,

≤ (θm−1 + θm−2 + · · · + θn )d(T x, x)

50 CHAPTER 2. EXISTENCE, UNIQUENESS AND INVARIANCE

by the contraction inequality

θn
≤ d(T x, x).
(1 − θ)

Now sending n → ∞ we see that the sequence is Cauchy and since the metric space
is complete, there exists an x in the space such that

lim T n x = x.
n→∞

Finally, x is a fixed point because

d(T x, x) = lim d(T n+1 x, T n x)

n→∞
= lim θn d(T x, x) = 0.
n→∞

QED
Chapter 3

The Geometry of Flows

3.1 Vector Fields and Flows

In this chapter we will tackle ODEs from a geometrical point of view. This is
a natural vantage point because an ODE is really a statement about the tangent
space to a manifold. The structure of the tangent space governs the geometry of
the manifold. First we define the objects that we will work with.

Definition 3.1 A function f : U → V , where U ⊂ Rn and V ⊂ Rm , is a Ck function,

or k-times differentiable, if its component functions y j are k-times differentiable

(y1 , y2 , . . . , ym ) = f (x1 , x2 , . . . , xn )

and the derivative ∂k y j /∂xlk , j, l ≤ k is continous. If k = ∞ then we say that f is

smooth.

Definition 3.2 A one to one map f : U → V , U ⊂ Rn , V ⊂ Rn , is a diffeomorphism

if both f and f −1 are Ck , k ≥ 1, functions.

Definition 3.3 A differentiable structure on a locally Euclidian space M is a col-

lection of open sets U j and coordinate maps ϕ j : U j → Rn , such that j ∈ J, J a set
of indicies,

1. ∪ j∈JU j = M

2. ϕ j o ϕ−1 k
k is C , for j, k ∈ J

51
52 CHAPTER 3. THE GEOMETRY OF FLOWS

3. The collection is maximal or if (U, ϕ) is any coordinate system then (U, ϕ) ⊂

{(U j , ϕ j ), j ∈ J} = F

Definition 3.4 An n-dimensional manifold of class k is a n-dimensional locally

Euclidian space M with a Ck differentiable structure F .

Example 3.1
The n-sphere is a n-dimensional manifold.
n
Sn = {x ∈ Rn+1 | ∑ x2j + (xn+1 − 1)2 = 1}
j=1

The differentiable structure consists of two open sets

U1 = Sn − (0, 0, · · · , 2), U2 = Sn − (0, 0, · · · , 0)

or the sphere with the two poles removed. The corresponding coordinate functions
are the stereographic projections from each punctured sphere ϕ1 = p1 , ϕ2 = p2 ,
see Figure 3.1. In physics the manifold M is the phase space of the physical system
and the motion on M is described by the ODE.

Definition 3.5 An orbit of a point x ∈ M is a map of a connected subset I of Rn

into M, φt (x) ∈ M, where t ∈ I.
We now specialize to autonomous systems to introduce some concepts that do
not always exist for non-autonomous systems.

Definition 3.6 gt is called a one-parameter group of transformation if it has the

following properties, for t, s ∈ I ⊂ R, and I is symmetric −I = I,
1. gt : M → M

2. gt gs = gt+s

3. g0 = identity
These properties imply that gt−1 = g−t . For t = T fixed, gT is nothing but the
time-T map, see Definition 2.7.

Definition 3.7 The one parameter group along with its manifold (M, {gt }) is
called the (phase) flow.
3.1. VECTOR FIELDS AND FLOWS 53

Figure 3.1: The two sphere with the two covering sets U1 and U2 and correspond-
ing steriographic projection p1 and p2 to R2 .
54 CHAPTER 3. THE GEOMETRY OF FLOWS

Lemma 3.1 gt is a one parameter group of diffeomorphism, if (M, {gt }) is a

flow and gt is differentiable.
Proof: That gt is a one parameter group of diffeomorphism implies that gt−1 =
g−t , now this in turn implies that gt−1 is differentiable since g−t is. Therefore, gt
is at least a C1 diffeomorphism. QED
Suppose that we start with some initial condition x ∈ M. x will move under
the flow and the solution to the initial value problem is its trajectory under the one
parameter group gt . The orbit of x becomes,
φt (x) = gt x, t ∈ I.
Thus it is the map of any connected time interval I ⊂ R into M, by gt . If I = R,
then the orbit is called a phase curve.
The above definitions are easily modified if we add the time axis to the phase
space to get the extended phase space M × R. Then (M × R, {gt }) is the extended
phase flow and the graph of the orbit (t, φt (x)) is called an integral curve. Notice
however that gt only acts on M not on M × R.

Exercise 3.1
1. Find the phase space of the harmonic oscillator
ẍ + ω2 x = 0
and the one parameter group of diffeomorphisms gt . Also find two different
type of orbits. Are these orbits phase curves?
2. Consider the nonlinear pendulum
ẍ + sin(x) = 0
Describe four different types of phase curves for the nonlinear pendulum.
We differentiate the orbit with respect to time to get the phase velocity
φ̇t (x) = v(x)
Since the orbit φt (x) = x(t) is just the solution the ODE, ẋ = f (x), with initial
condition x(0) = x, we conclude that the phase velocity is just the vector field
v(x) = f (x)
This is where the name vector field originates, v is the vector field on the phase
space M that determines the flow.
3.1. VECTOR FIELDS AND FLOWS 55

Definition 3.8 The (phase) velocity of a flow

ẋ = φ̇t (x) = v(x) = f (x)
is the vector determining the flow at the point x ∈ M. A point x̄ where
v(x̄) = 0
is called a singular point of the vector field.
It is clear from the definition that if φt (x) is Ck then v(x) is Ck−1 . By the
Picard Existence Theorem 2.1, we know that if v(x) is Ck−1 then φt (x) is Ck . In
the extended phase space v defines a direction field and a curve is an integral curve
if and only if it is always tangent to the direction field, see Figure 3.1.

Example 3.2
1. ẋ = kx
2. ẋ = sin(x)
1. The phase space is M = R, the one-parameter group of diffeomorphism is,
gt = ekt
and the orbit is
φt (x) = ekt x
x̄ = 0 is the single stationary solution and singular point of the vector field.
The flow is away from 0, so 0 is a unstable stationary solution, see Figure
3.1.
2. The phase space is M = R, but gt is not a simple function, see the Exercise
3.2. However, gt and φt (x) exist and can be worked out. The stationary
solutions and singular point of the vector field are x = nπ, n ∈ Z. The even
multiples of π are unstable and the odd multiples are stable, see Figure 3.1.

Exercise 3.2
1. Find the second semi-group in Example 3.2
2. Find gt and φt (x) for Example 2.
3. Show that a one parameter group of transformation may not always exist or
only exists for a finite time.
56 CHAPTER 3. THE GEOMETRY OF FLOWS

Figure 3.2: The direction vector field tangent to the integral curve of the orbit.
3.1. VECTOR FIELDS AND FLOWS 57

Figure 3.3: The extended phase space and the phase space of the ODEs 1 and 2,
in Example 3.2, respectively.
58 CHAPTER 3. THE GEOMETRY OF FLOWS

3.2 The Tangent Space

We now consider the phase velocity from a slightly different point of view, namely
as represented by its coordinates in Euclidian space. Let (U, ϕ) be a member of
the differentiable structure on the manifold M, such that
x ∈ U and ϕ = (x1 , x2 , . . . , xn )
the x j s being coordinate functions. v = (v1 , v2 , . . . , vn ) where
d
x j o φt (x)|t=0 , j = 1, . . . , n
v j (x) = (3.1)
dt
Moreover, we define two orbits (or just curves) to be tangent at x if their difference
is order t 2
kφt1 (x) − φt2 (x)k = O(t 2 )
A change of coordinates is called admissible if it is a diffeomorphism. This notion
of a class of admissible changes of coordinates permits us to define a tangent
vector without a reference to a coordinate system.

Definition 3.9 The velocity vector to an orbit at a point x ∈ M

φ̇t (x) = v(x)
is the equivalence class of tangents to the equivalence class of orbits at x. The
tangent space at x : T Mx is the equivalence class of all tangent vectors to orbits
starting at x.
It is clear that two orbits are equivalent if there is a diffeomorphism taking one
into the other, also if
y:U →V
is a diffeomorpishm, then the Jacobian
Dx y : TUx → TVy
is the corresponding map of the tangent spaces and this map is linear. We simply
differentiate
φt2 (y(x)) = y(φt1 (x))
with respect to t, to get that
v2 (y(x)) = φ̇t2 = Dx y φ̇t1 = Dx y v1 (x)
by the chain rule. Moreover, it is easy to prove that the coordinate map (3.1) of v
is one to one.
3.2. THE TANGENT SPACE 59

Lemma 3.2 The coordinate map

v : T Mx → Rn

is an isomorphism.

Proof: We have to show that if two tangent vectors v1 = v2 are the same then the
corresponding orbits are tangent. Since

v1 − v2 = 0

Dx y (φ̇t1 − φ̇t2 ) = 0.
However, since Dx y is invertible it follows that

φ̇t1 − φ̇t2 = 0.

This implies that

kφt1 (x) − φt2 (x)k = O(t 2 ).
QED

A sufficient condition for a differentiable map to be a diffeomorphism is given

by the next theorem.

Theorem 3.1 The Inverse Function Theorem Let g : U → V be a C1 map from Rn

to Rn and suppose that detDx g 6= 0, at x ∈ U. This implies that

Dx g : TUx → TVy

is an isomorphism and then there exists a neighborhood W ⊂ U of x, such that

g : W → g(W )

and g restricted to W is a diffeomorphism.

Proof: Since Dx g is an isomorphism dimU = dimV . Let x and y be coordinates

on U and V respectively. Define

F(x, y) = y − g(x)

such that
F(x0 , y0 ) = y0 − g(x0 ) = 0.
60 CHAPTER 3. THE GEOMETRY OF FLOWS

Then
detDx F = −detDx g 6= 0,
at x = x0 , and by the implicit function theorem there exists a small neighborhood
E ⊂ V of y0 and a function

h : E → h(E) = W

such that
F(h(y), y) = 0
for y ∈ E. Moreover, h is unique and differentiable as often as g. Clearly

h = g−1 ∈ C1 ,

therefore g is a diffeomorphism. QED

3.3 Flow Equivalance

The nonautonomous ODE
ẋ = f (x,t)
can be made autonomous by introducing a new variable θ = t, then θ̇ = 1 and the
above system is equivalent to the autonomous system

ẋ = f (x, θ)
θ̇ = 1

Thus our definition of the tangent vector and tangent space hold in the extended
phase space M × R with coordinates (x, θ). However, frequently this is not the
most desirable result, we would like objects such as the semi-group to be defined
for the phase space M itself and this is not always possible to do. But the addition
of θ as above can be a very useful tool and it is just what we need for the next
theorem. It is called the Rectification Theorem and it says that in a neighborhood
of a nonsingular point any flow can be mapped onto straight linear flow.

Theorem 3.2 The Rectification Theorem.

Let (x0 , θ0 ) be a nonsingular point of the vector field

ẋ = f (x, θ)
θ̇ = 1
3.3. FLOW EQUIVALANCE 61

then there exists a neighboorhood U of x and a diffeomorphism g : U → V ⊂ Rn+1

such that the ODE becomes

ẏ = 0
ż = 1

where (y, z) = g(x, θ). Moreover, if f is Ck then g is also Ck .

Proof: We define a map h from the line (y, z) to the orbit (φt (x), θ). If we can
show that h is a diffeomorphism, then g = h−1 is the desired map. Define

(φt (x), θ) = h(y, z)

where y = φt (x), a constant in local coordinates, and z = θ. h is clearly differen-

tiable with respect to y and z and
 
I φ̇t
D(y,z) h =
0 1

at x = φt0 , θ = θ0 . It is easy to see that D(y,z) h does what it is supposed to do. It

maps the vector field (0, 1) onto (v, 1), where v = φ̇t ,
    
I v 0 v
=
0 1 1 1

Moreover
detD(y,z) h = 1 6= 0
and by the inverse function theorem 3.1 h is a diffeomorphism. If v = f is Ck−1 ,
then h is Ck and g is also Ck by the inverse function theorem. QED

It is helpful to view the proof of the Rectification Theorem geometrically.

D(y,z) h maps Rn × R onto Rn × V , where V = span(v, 1). However, (v, 1) is
transverse to Rn , see Figure 3.3, so both spaces are n + 1 dimensional. Now
kernalD(y,z) h is empty so the map is an isomorphism. The inverse of D(y,z) h is
D(x,θ) g, the derivative of g.

Exercise 3.3 Prove the Picard Theorem 2.1 (away from a singular point) using
the Rectification Theorem.
62 CHAPTER 3. THE GEOMETRY OF FLOWS

Figure 3.4: The two vector fields in the Rectification Theorem and the map be-
tween them.
3.3. FLOW EQUIVALANCE 63

Definition 3.10 Flow Equivalence

Two flows are conjugate (equivalent) if there exists a one to one map g between
corresponding orbits or
φt2 o g = g o φt1
The flows are
1. linearly conjugate if g is a linear map, then g ∈ C∞ ,
2. differentiably conjugate if g is a diffeomorphism, g ∈ Ck , k ≥ 1,
3. topologically conjugate if g is a homoeomorphism g ∈ C0 .

Lemma 3.3 Two linear systems are linearly conjugate if and only if they have the
same eigenvalues with the same algebraic and geometric multiplicity.
Proof: Consider the systems
ẋ = Ax and ẏ = By
they are linearly conjugate if
φt2 (h(x)) = hφt1 (x)
where h is an n × n invertible matrix. Now
φ̇t2 = hφ̇t1 = hAφt1
= hAh−1 hφt1 = hAh−1 φt2
Comparing
φ̇t2 = hAh−1 φt2 and ẏ = By
we conclude that A and B are similar matricies
B = hAh−1
Conversely, if A and B are similar matricies then the equations
φ̇t2 = hAh−1 φt2 and ẏ = By
can both be solved with the same initial data to give the solution
φt2 (hx) = hφt1 (x)
by the uniqueness of the solution to the initial value problem. This shows that the
two orbits are linearly conjugate. QED
64 CHAPTER 3. THE GEOMETRY OF FLOWS

Theorem 3.3 Two linear systems are differentiably conjugate if and only if they
are linearly conjugate.

Proof: Consider the systems

ẋ = Ax and ẏ = By

If the orbits are differentiably conjugate by Lemma 3.3 is suffices to show that A
and B are similar matricies. Suppose that

φt2 (g(x)) = g(φt1 (x)) (3.2)

where g is at least in C1 . We expand the orbits in t:

φt1 (x) = φt=0

1 1
(x) + φ̇t=0 (x)t + O(t 2 ) = x + Axt + O(t 2 )

Similarly,
φt2 (x) = y + Byt + O(t 2 )
We substitute these expansions into Equation 3.2 and differentiate with respect to
the initial data x. This can be done because g is a diffeormophism and the orbits
depend continuously on their initial data. The differentiation of

g(x) + Bg(x)t + O(t 2 ) = g(x + Axt + O(t 2 ))

gives
(I + Bt + O(t 2 ))g0 = g0 (I + At + O(t 2 ))
where g0 is the Jacobian of g. This shows that

Bg0 = g0 A

and the matricies are similar. The converse is immediate because a linear conju-
gacy is a C∞ diffeomorphism. QED

Remark 3.1 Theorem 3.3 shows that differentiable conjugacy is too strict an
equivalence relation at a singular point of a vector field. We must settle for topo-
logical conjugacy.
We will show below that generically a nonlinear flow is topologically con-
jugate to its linearization. The topological conjugacy classes are given by the
invariant manifold theorems. Namely, the stable and unstable manifold theorem
and the center manifold theorem.
Chapter 4

Invariant Manifolds

The linear stable and unstable manifolds Es and Eu of a hyperbolic stationary

solution, consisting of the eigenvectors of the stable and unstable eigenvalues re-
spectively, are the tangent spaces of their nonlinear counterparts, at the origin. The
existence of the nonlinear stable and unstable manifolds, see Figure 4 is given by
the following theorem.

Theorem 4.1 The Invariant Manifold Theorem

Consider the equation
ẋ = Ax + g(x), x ∈ Rn
where A has n eigenvalues λ j , and Reλ j 6= 0; g(x) is Ck in a neighborhood of
x = 0 and
kg(x)k
lim = 0,
kxk→0 kxk

then there exists a neighborhood U of the origin, a Ck manifold W s and a Ck

function hs : Πs (U) → Eu , where Πs (U) is the projection of U onto Es , such that
∂hs
1. hs (0) = 0 and ∂xs (0) = 0, W s is the graph of hs

2. x(t0 ) ∈ W s =⇒ x(t) ∈ W s , ∀ t ≥ t0

3. x(t0 ) 6= W s =⇒ ∃ δ > 0, t1 ≥ t0 , kx(t)k > δ, ∀ t ≥ t1

There also exists a Ck manifold W u and a Ck function hu : Πu (U) −→ Es , where

Πu (U) is the projection of U onto Eu , such that,
∂hu
1. hu (0) = 0 and ∂xu (0) = 0, and W u is the graph of hu

65
66 CHAPTER 4. INVARIANT MANIFOLDS

Eu

u
W

s
W

Es

Figure 4.1: The Invariant Manifolds around a Hyperbolic Stationary Orbit.

2. x(t0 ) ∈ W u =⇒ x(t) ∈ W u , ∀ t ≤ t0
3. x(t0 ) 6= W u =⇒ ∃ δ > 0, t1 ≤ t0 , kx(t)k > δ, ∀ t ≤ t1

Proof: By the statement of the theorem we both have to prove the existence of the
functions hs and hu whose graphs are the manifolds W s and W u and the existence
of the solutions to the ODE xs (t) ∈ W s and xu (t) ∈ W u .
If Es is the linear stable manifold of the linearized system ẏ = Ay then A leaves
Es invariant, AEs ⊂ Es , and if we define the exponential of a matrix to be the power
series
A2t 2 Ant n
etA = I + At + +···+ +··· ,
2 n
then the exponential also leaves Es invariant, etA Es ⊂ Es . Moreover, we have an
estimate
ketA ys k ≤ Cs e−ρt kys k, t ≥ 0 (4.1)
and the same remarks apply to Eu so etA Eu ⊂ Eu and
ketA yu k ≤ Cu eσt kyu k, t ≤ 0 (4.2)
Cs , and Cu are constants. −ρ and σ are respectively the smallest (in absolute value)
negative and positive real parts of the eigenvalues of A.
 67

We now express the ODE as an integral equation

Z t
(t−t0 )A
x(t) = e x(t0 ) + e(t−τ)A g(x(τ))dτ.
t0

Since Rn is a direct sum Rn = Es ⊕Eu we can split x = xs +xu and write an integral
equation for xs and xu separately
Z t
(t−t0 )A
xs (t) = e xs (t0 ) + e(t−τ)A g(x(τ))s dτ
t0

and Z t
(t−t0 )A
xu (t) = e xu (t0 ) + e(t−τ)A g(x(τ))u dτ.
t0
We have used here that both Es and Eu are invariant under the flow. Next we
look for a solution that lies in W s and therefore remains in a neighborhood of the
origin for all t ≥ 0. If we consider the second equation and notice that xu picks up
the positive eigenvalues of A, we realize that the only way we are going to get a
solution that stays in a neighborhood of the origin is to send t0 → ∞. Then
0
lim ke−t0 A xu (t0 )k ≤ lim Cu0 e−σ t0 = 0,
t0 →∞ t0 →∞

assuming xu (t) stays bounded, where σ0 is now the smallest positive real part of
the eigenvalues of A. This gives
Z t
xu (t) = e(t−τ)A g(x(τ))u dτ
∞

and adding this to the equation for xs above, with t0 = 0, gives

Z t Z ∞
x(t) = etA xs (0) + e(t−τ)A g(x(τ))s dτ − e(t−τ)A g(x(τ))u dτ. (4.3)
0 t

Now let xs (0) = xs0 ∈ Es and recall some spaces that we already encountered
in the proof of Picard’s theorem. First let

C0 ([0, ∞); Rn )

be the space of continuous functions from the half-line R+ = [0, ∞) into Rn . Each
point in C0 ([0, ∞); Rn ) is a n-vector valued function x(t) ∈ Rn , for t fixed. How-
ever, C0 ([0, ∞); Rn ) is not a Banach space under the sup norm. It is a Fréchet
68 CHAPTER 4. INVARIANT MANIFOLDS

space or a complete metric space under a sequence of quasi-norms but this is not
what we need. We have to introduce a weight eµt where µ = min(ρ, σ) to make
Cµ0 ([0, ∞); Rn ) into a complete normed linear space, or in other words a Banach
space, under the norm
kx(t)kµ = sup eµt kx(t)k.
t∈[0,∞)

We have forced the function in Cµ0 ([0, ∞); Rn ) to decay exponentially in t by in-
serting the weight eµt and this makes the space a Banach space. We will restrict
the space further by making the functions lie in a small ball in Rn and define the
space on which we will prove the contraction to be
0
Cµ,δ ([0, ∞); Rn ) = {x(t) ∈ Cµ0 | kx(t)k ≤ δ}.

This restriction partially destroys the Banach space structure because the sum of
0 in no longer lies in this ball, x + y ∈ C0 , so it is not a linear
two vectors x, y ∈ Cµ,δ µ,δ
space. However, it is easy to check that this space is a complete metric space and
this is all we need for an application of the Contraction Mapping Principle 2.10.
Now we use the integral equation (4.3) to define a map
Z t Z ∞
(t−τ)A
tA
F (x(t)) = e xs (0) + e g(x(τ))s dτ − e(t−τ)A g(x(τ))u dτ. (4.4)
0 t

0 , if kx0 k < γ, γ small. This last inequality

Then F (x(t)) is a contraction on Cµ,δ s
defines the small neighborhood U of the origin, in the statement of the theorem.
First we must show that F maps Cµ,δ 0 into itself. We multiply the equation

(4.3) by eµt and apply the triangle inequality and the two inequalities (4.1) and
(4.2) above, and the hypothesis on g, in the first step; since lim|x|→0 kg(x)k
kxk = 0, so
that there exists ε > 0 such that kDg(x)k ≤ εkxk,
Z t
−(ρ−µ)t
µt
e kx(t)k ≤ Cs e kxs (0)k +Cs ε e−(ρ−µ)(t−τ) eµτ kx(τ)kdτ
Z ∞ 0

+ Cu ε e(σ+µ)(t−τ) eµτ kx(τ)kdτ

t
 
Cs Cu
≤ Cs kxs (0)k + ε + kx(t)kµ
(ρ − µ) (σ + µ)
 
Cs Cu
≤ Cs γ + ε + δ
(ρ − µ) (σ + µ)
≤ δ,
 69

for kxs (0)k ≤ γ and ε small enough.

Secondly we show the F is a contraction. We need to show that there exists
0 < θ < 1 such that

kF(x1 (t)) − F(x2 (t))k ≤ θkx1 (t) − x2 (t)k,

where the map F is defined by the integral equation. Consider

eµt [F(x1 (t)) − F(x2 (t))] =

Z t
µt tA
e e 0
(x1,s 0
− x2,s ) + eµt e(t−τ)A [g(x1 (τ))s − g(x2 (τ))s ]dτ
Z ∞ 0

−eµt e(t−τ)A [g(x1 (τ))u − g(x2 (τ))u ]dτ

t
Z t
−(ρ−µ)t
≤ e 0
kx1,s 0
− x2,s k + εCs e−(ρ−µ)(t−τ) eµτ kx1 (τ) − x2 (τ)kdτ
Z ∞ 0

+εCu e(σ+µ)(t−τ) eµτ kx1 (τ) − x2 (τ)kdτ,

t

by the estimates above and the hypothesis on g. Estimating as above, we get

≤ e−(ρ−µ)t kx1,s
0 0
− x2,s k
 
Cs Cu
+ ε +ε sup eµt kx1 (t) − x2 (t)k.
(ρ − µ) (σ + µ) [0,∞)

0 = x0 we get a unique fixed point for

In particular, for x1,s 2,s
 
Cs Cu
ε + = θ < 1.
(ρ − µ) (σ + µ)

We have shown that the integral equation (4.3) has a unique solution x(t), but
this gives us a function hs : Es → Eu . Namely, x(t) = (xs , xu )(t) where
Z t
(t−t0 )A
xs (t) = e xs (t0 ) + e(t−τ) gs (x(τ))dτ
t0

and Z ∞
xu (t) = − e(t−τ)A gu (x(τ))dτ.
t
70 CHAPTER 4. INVARIANT MANIFOLDS

At t = t0 , we get xs (t0 ) = xs and

Z ∞
s
h (xs ) = xu (t0 ) = − e(t−τ)A gu (x(τ)) dτ.
t0

This is the desired function. The dependance of hs on xs enters throught the de-
pendance on x(τ) that depends on the initial data xs (t0 ) = xs .
We will now prove that the function hs has the properties (1) - (3). To prove
(1) we consider the integral equations above that xs (t) satisfies,
Z t
xs (t) = etA xs (0) + e(t−τ)A g(x(τ))s dτ,
0

where we have chosen t0 = 0 for convenience. We get using the triangle inequality
for the norm and taking the limit as t → ∞
Z t
−ρt −µt
lim kxs (t)k ≤ lim Cs e kxs (0)k + lim Cs εe e−(ρ−µ)(t−τ) eµτ kx(τ)kdτ
t→∞ t→∞ t→∞ 0
Cs ε
≤ Cs lim e−ρt kxs (0)k + lim e−µt kx(t)kµ = 0.
t→∞ (ρ − µ) t→∞

This says that

lim xs (t) = 0
t→∞
and consequently
hs (0) = 0
because
Z ∞
khs (0)k = lim kxu (t)k ≤ lim Cu εe−µt e(σ+µ)(t−τ) eµτ kx(τ)kdτ
t→∞ t→∞ t
Cu ε
≤ lim e−µt kx(t)kµ = 0.
(σ + µ) t→∞
Then we differentiate the equation for h
Z ∞
s
Dxs h = − e(t−τ)A Dx g(x(τ))u Dxs x(τ)dτ
t

This integral converges uniformly, since Dxs x is bounded by the Continous De-
pendence on Parameters Theorem and

Dxs hs (0) = lim Dxs xu (t) = 0

t→∞
 71

by the estimate
Z ∞
−µt
s
kDxs h (0)k = lim Dxs xu (t) ≤ lim Cu e e(σ+µ)(t−τ) eµτ kx(τ)kkDxs x(τ)kdτ
t→∞ t→∞ t
C
≤ lim e−µt kx(t)kµ = 0.
(σ + µ) t→∞
To prove (2) we consider the integral equation (4.3). If x(t0 ) ∈ W s then
Z ∞
xu (t0 ) = − e(t−τ)A g(x(τ))u dτ,
t0

and xs (t0 ) = xs . Solving the equations for xu (t) and xs (t) with this initial data, we
get Z t
xs (t) = e(t−t0 )A xs + e(t−τ)A g(x(τ))s dτ (4.5)
t0
Z ∞
xu (t) = − e(t−τ)A g(x(τ))u dτ. (4.6)
t
However, adding the two equations (4.5) and (4.6) and setting t0 = 0, gives the
equation (4.3) that characterizes W s so x(t) ∈ W s for t ≥ 0.
The property (3) is proven by an estimate. If we write the original ODE in
integral form then
Z t Z t
(t−t0 )A (t−τ)A (t−t0 )A
x(t) = e xu (t0 )+ e g(x(τ))u dτ+e xs (t0 )+ e(t−τ)A g(x(τ))s dτ.
t0 t0

We let σ0 be the smallest positive real part of the eigenvalues of A, σ is the largest
positive real part of the eigenvalues of A as above. Now by the triangle inequality
and the estimates (4.1) and (4.2)
Z t
σ0 (t−t0 ) −µt
kx(t)k ≥ Cu e kxu (t0 )k −Cu εe e(σ+µ)(t−τ) eµτ kx(τ)k)dτ
t0
Z t
−ρ(t−t0 ) −µt
− Cs e kxs (t0 )k −Cs εe e−(ρ−µ)(t−τ) eµτ kx(τ))kdτ
t0
0 ε
≥ Cu (eσ (t−t0 ) kxu (t0 )k − e−µt kx(t)kµ ) −Cs (e−ρ(t−t0 ) kxs (t0 )k
(σ + µ)
ε
+ e−µt kx(t)kµ ) ≥ δ,
(ρ − µ)
for t − t0 sufficiently large. This shows that if xu (t0 ) 6= 0 so x(t0 ) 6∈ W s then x(t)
must leave a neighborhood of the origin at some time t1 ≥ t0 and stay out for t ≥ t1 .
QED
72 CHAPTER 4. INVARIANT MANIFOLDS

How do we compute the stable and the unstable manifolds? We derive an equation
for hs (and hu ) and approximate these function by quadradic and higher Taylor
polynomials. The equations can be written as

ẋ = Ax + f (x, y), x ∈ E s
ẏ = By + g(x, y), y ∈ E u .

Now we substitute in y = hs , by the chain rule

Dx hs ẋ = By + g(x, hs )

or
Dh(Ax + f (x, h)) = Bh(x) + g(x, h)
where we dropped the superscript and all functions are functions of x only. This
is the equation satisfied by hs (x), similarly hu (y) satisfies the equation

Dh(By + g(h, y) = Ah + f (h, y).

Example 4.1        2
d x1 −1 0 x1 −x2
= +
dt x2 0 1 x2 x12

ẋ = −x − y2 A = −1
ẏ = y + x2 B=1

We want to find the function hs : Es → Eu whose graph is the stable manifold

W s . hs satisfies the equation

Dh(Ax + f (x, h)) = Bh + g(x, h)

or in our case
h0 (−x − h2 ) = h + x2 .
We look for a solution that is a power series in x

h(x) = ax2 + bx3 + · · ·

where a and b are unknown coefficients to be determined,

h0 (x) = 2ax + 3bx2 + · · · .

 73

Substituting h and h0 into the equation above we obtain

(2ax + 3bx2 + · · · )(−x − (ax2 + bx3 + · · · )2 ) = ax2 + bx3 + · · · + x2

or
−2ax2 − 3bx3 + · · · = (a + 1)x2 + bx3 + · · · .
Equating coefficients gives

−2a = a + 1, a = −1/3
−3b = b, b = 0.

Thus
−x2
hs (x) = + O(x4 ).
3
Exercise 4.1

1. Find the linear and nonlinear stable and unstable manifolds of the hyper-
bolic stationary solutions of the Duffing equations,

ẍ ± (x − x3 ) = 0,

and draw them. What difference do the ± signatures make?

2. Find the stable and unstable manifolds of the hyperbolic stationary solution
at the origin for the system
       3
d x 1 0 x −y
= +
dt y 0 −1 y x3
74 CHAPTER 4. INVARIANT MANIFOLDS
Chapter 5

Chaotic Dynamics

5.1 Maps and Diffeomorphisms

We have already discussed in Definition 2.7 the time-T map of a flow which is a
Ck+1 diffeomorphism if the flow is invertible and the vector field a Ck function.
Now we consider general Ck diffeomorphisms and their connections to flows.

Definition 5.1 A map

xm+1 = f (xm ), xm ∈ Rn , m ∈ Z (5.1)

is a Ck diffeomorphism if f : Rn −→ Rn is a Ck diffeomorphism.
The orbit of the map {xm } ⊂ Rn is now a sequence generated by composing
the function f with itself. We will use the shorthand

f m (x) = f ◦ f ◦ · · · ◦ f (m times).

Maps have their local existence theory analogous to the ODE theory in Chapter I
and a linear theory analogous to the theory of linear ODE’s. We will not review
this theory here. A nice account of the theory of linear ODE’s can be found in
Perko [17] Chapter 1.
We will use diffeomorphism as a tool to understand flows (one can also use
flows as a tool to understand diffeomorphisms) and in particular, we are interested
in the analog for maps of stationary solutions of ODE’s and their stability.

Definition 5.2 x is called a fixed point of a map (5.1) if f (x) = x. x∗ is a periodic

point of period m if f m+1 (x∗ ) = x∗ .

75
76 CHAPTER 5. CHAOTIC DYNAMICS

Figure 5.1: The Quadradic Map

5.1. MAPS AND DIFFEOMORPHISMS 77

Example 5.1 Consider the quadratic map

xm+1 = f (xm )

where
f (x) = 1 − µx2 ,
see Figure 5.1. This map has two fixed points where the graph of f meets the line
y = x and since one of these points has slope less than one, it is stable. The other
has slope greater than one and is unstable, see below.

Definition 5.3 An orbit {xm } of a map (5.1) is stable if for ε > 0, there exists
δ > 0, such that

kx0 − y0 k < δ ⇒ kxm − ym k < ε, m ≥ 1.

{xm } is asymptotically stable if

lim ym = xm .
m→∞

Stability can be proven by a Lyapunov function but we are mostly interested in

fixed points and periodic orbits which are hyperbolic. Consider the map (5.1)
linearized about a fixed point x̄,

xm+1 − x̄ = D f (x̄)(xm − x̄)

then the stability of x̄ is given by the following lemma.

Lemma 5.1 If |λ| < 1 for all λ ∈ σ(D f (x)) then x is (asymptotically) stable. If
|λ| > 1 for some λ ∈ σ(D f (x)) then x is unstable.
Here σ(D f (x)) denotes the spectrum of f linearized about the fixed point x.
Proof: By the Mean-Value Theorem
Z 1
xm − ym = D f (sxm−1 + (1 − s)ym−1 ) ds · (xm−1 − ym−1 ).
0

Thus if xm = x and kx − ym k < δ, then kx − ym k ≤ (max |λ| + η)kx − ym−1 k, where

η can be made arbitrarily small by making δ sufficiently small. Now iteration
gives kx − ym k ≤ (max |λ| + η)m kx0 − y0 k. If max |λ| < 1 then

lim ym = x.
m→∞
78 CHAPTER 5. CHAOTIC DYNAMICS

If there exists an eigenvalue |λ| > 1 then we can pick x−y0 to be the corresponding
eigenvector. Then
Z 1
x − y1 = D f (sx + (1 − s)y0 ) ds · (x − y0 ),
0

by the mean value theorem and

kx − y1 k ≥ (|λ| − η)kx − y0 k,

where η is small. An iteration gives

kx − ym k ≥ (|λ| − η∗ )m kx − y0 k,

where η∗ is still small. This shows that {ym } diverges from x. QED
If x∗ is a periodic orbit of period m then we get the criteria,

Corollary 5.1 If |λ| < 1 for all λ ∈ σ(D f m (x∗ )) then x∗ is (asymptotically) stable.
If |λ| > 1 for some λ ∈ σ(D f m (x∗ )) then x∗ is unstable.
Invariant sets for maps are defined in the same way as those for flows in Chap-
ter 1. As for ODE’s we are interested in conjugacy classes of maps.

Definition 5.4 Two maps f and g are topologically conjugate if

h◦g = f ◦h

where h is a homeomorphism.

Lemma 5.2 Suppose f : R → R is a diffeomorphism and D f (x) > 0 for some

x ∈ R. Then the map xm+1 = f (xm ) is topologically conjugate to the time-1 (T = 1)
map of the flow ϕt defined by ẋ = f (x) − x.

Proof: Since f is a diffeomorphism and D f (x) > 0 for some x, it follows that
D f (x) > 0 for all x ∈ R. Consider the graph of f and its intersections with the line
y = x in Figure 5.1. The fixed points of the map xi = f (xi ) are clearly equal to the
stationary solutions of the ODE

ẋi = f (xi ) − xi = 0.

Moreover the orientation or the direction in which the points map or flow is the
same for the map and the flow in each subinterval [xi , xi+1 ]. This is because sign
5.1. MAPS AND DIFFEOMORPHISMS 79

Figure 5.2: The graph of the function f (x) and the line y = x.
80 CHAPTER 5. CHAOTIC DYNAMICS

[xm+1 − xm ] = sign [ f (xm ) − xm ] = sign ẋ. Now let x0 ∈ (xi , xi+1 ) be any point and
define the intervals
f : [Pm , Pm+1 ] −→ [Pm+1 , Pm+2 ]
where Pn = f n (x0 ) and
ϕ1 : [Qm , Qm+1 ] −→ [Qm+1 , Qm+2 ]
where Qm = ϕm (x0 ). Both f and ϕ1 move the intervals in the same direction.
We want to construct a homeomorphism taking the orbits of ϕ onto the orbits of
f . First we construct a homeomorphism taking [Q0 , Q1 ] onto [P0 , P1 ]. There are
many ways of doing this but we will pick the simplest and let
 
f (x0 ) − x0
h0 (y) = x0 + [y − y0 ].
ϕ1 (y0 ) − y0
Now clearly h(y0 ) = P0 and h(y1 ) = h(ϕ(y0 )) = f (x0 ) = P1 , so h[Q0 ] = P0 and
h[Q1 ] = P1 . Moreover, h is monotone (linear) and thus maps [Q0 , Q1 ] onto [P0 , P1 ]
in a one to one and invertible fashion. For the mth intervals we simply define
hm (y) = f m ◦ h0 ◦ ϕ−m (y).
In other words hm pulls [Qm , Qm+1 ] back to [Q0 , Q1 ] by ϕ−m , maps it to [P0 , P1 ] by
h0 and hm moves it forward to [Pm , Pm+1 ] by f m . The desired homeomorphism is
then 
 xi for y = xi
h(y) = hm (y) for y ∈ (Qm , Qm+1 )
xi+1 for y = xi+1 .


Finally, we check that h has the right property,

h ◦ ϕ1 = hm ◦ ϕ1 = f m ◦ h0 ◦ ϕ−m ◦ ϕ1
= f ◦ f m−1 ◦ h ◦ ϕ−(m−1) = f ◦ h.
QED
If two maps are topologically conjugate they are said to be of the same topological
type. One has to construct a homeomorphism between them to show that two
maps are conjugate. It is on the other hand frequently easier to show that two
maps are not topologically conjugate for example if they have a different number
of fixed point (and periodic orbits) then they must belong to different conjugacy
classes. The argument is simply that otherwise a non-trivial orbit gets mapped
onto a fixed point by a homeomorphism and this is impossible.
5.2. CLASSIFICATION OF FLOWS AND MAPS 81

5.2 Classification of Flows and Maps

We can now give a complete local classification of flows and maps up to topo-
logical conjugacy. By the Hartmann-Grobmann Theorem, see Hartman [12], a
nonlinear system
ẋ = f (x)
is topologically conjugate to its linearization

ẏ = D f (x)y, y = x − x,

at a hyperbolic stationary solution x. Then the Rectification Theorem 3.2, see

Arnold [2], says that the flow on the linear unstable manifold E u and the linear
stable manifold E s are topologically equivalent to the flows of

ẋu = xu , and ẋs = −xs

respectively, where xu = pu ◦ x and xs = ps ◦ x denote the projections onto E u and

E s . This proves the following theorem.

Theorem 5.1 The flow of the system

ẋ = f (x)

at a hyperbolic stationary solution x is topologically conjugate to the flow of

 
Iu 0
ẋ = x
0 −Is

in a neighborhood of the origin, where rank Iu = dim E u and rank Is = dim E s .

This means that we only have to count the different configurations with respect to
the imaginary axis that n eigenvalues can have, to make a complete list of the local
topological conjugacy classes at a hyperbolic stationary solutions in n dimensions.
This count is easy, just place all n eigenvalues in the right half plane and then move
one at a time over to the left half plane to get n + 1 configurations. Thus there are
n + 1 different local conjugacy classes of flows at a hyperbolic stationary solution.
The discussion is similar for maps. The map

xm+1 = f (xm ), m ∈ Z,
82 CHAPTER 5. CHAOTIC DYNAMICS

is topologically conjugate to its linearization

ym+1 = D f (x)ym , ym = xm − x
at a hyperbolic fixed point. Now if we decompose D f (x) into two matricies, U
with eigenvalues inside and S with eigenvalues outside the unit circle, then U
(or S) is orientation-preserving if detU > 0 and orientation-reversing if detU <
0. Recall that the eigenvalues come in conjugate pairs because f (xm ) is real.
Then the Hartmann-Grobmann Theorem, see Hartman [12], proves the following
theorem.

Theorem 5.2 The nonlinear map

xm+1 = f (xm ), m ∈ Z
is topologically conjugate at a hyperbolic fixed point to the map
 
U 0
ym+1 = y
0 S m
where λ ∈ σ(U) =⇒ |λ| > 1, λ ∈ σ(S) =⇒ |λ| < 1, rank U = dim E u , rank S =
dim E s and U and S are either orientation-preserving or orientation-reversing.
Now we count how many local conjugacy classes there are for maps. Two
maps are topologically conjugate at a hyperbolic fixed point if and only if rank
U1 = rank U2 , rank S1 = rank S2 and the orientation of U1 and U2 and respectively
S1 and S2 are the same. This means that we just have to count in how many
ways we can configure n eigenvalues with respect to the unit circle. That is n + 1
configurations as above. However, only n − 1 of those have 4 orientation namely
++, +−, −+ and −−, where the first signature refers to detU and the second
to det S. Two configurations have only two possible signature, the eigenvalues are
all inside the unit circle, U = 0, and all outside S = 0. This gives
4(n − 1) + 2 · 2 = 4n
local equivalence classes for maps in n dimensions. We now consider the one-
dimensional case to explain and then prove the converse of the statement above.
There are four possible positions of the eigenvalues with respect to the unit circle
and positive and negative real axis. These are illustrated in Figure 5.2, the first
two corresponding maps, on Figure 5.2, are orientation preserving the last two
are orientation reversing. Now if the signatures of detU1 and detU2 are different,
then U1 has a different number of (generalized) eigenvectors along which the map
looks like case 1 (and case 4) and they are not topologically conjugate.
5.2. CLASSIFICATION OF FLOWS AND MAPS 83

Figure 5.3: The positions of the eigenvalues and the corresponding motion of
the iterates of a one-dimensional map in phase space. The first two maps are
orientation preserving and the last two orientation reversing.
84 CHAPTER 5. CHAOTIC DYNAMICS

Exercise 5.1 Draw the eigenvalue configurations and map phase portrait for the
8 two-dimensional cases.

5.3 Horseshoe Maps and Symbolic Dynamics

The canonial example of a diffeomorphism with a complex invariant set is the
horseshoe map invented by Steven Smale. Consider the square S = {(x, y)|
|x|, |y| ≤ 1} and the map that consist of the following two acts:

1. Stretch S in the x- direction and compress it in the y- direction such as to

map S onto the horizontal rectangle

R = {(x, y)| |x| ≤ 5, |y| ≤ 1/5}.

2. Bend R into a horseshoe and place it on top of S so that the intersection of

S and the horseshoe consists of two horizontal rectangles H0 and H1 , see
Figure 2.

Heuristically, the stretching and compressing accounts for the hyperbolic part of
the map and the bending makes it nonlinear. Notice that the preimages of H0 and
H1 are vertical strips V0 and V1 and now we define the horseshoe map on S,

f (S) ∩ S = H0 ∪ H1

to be those two steps. In particular,

f (V j ) = H j , j = 0, 1.

The precise form of the map on SV0 ∪V1 is not crucial, it is linear on V j , j = 0, 1.
We take it to be as in Figure 2 and this becomes important when we embed the
map later in the sphere. The image of H j , j = 0, 1, in S, consists of four horizontal
strips H jk , j, k = 0, 1, see Figure 5.3. In general, the image of H n = f (H n−1 ) ∩
S, H 0 = H0 ∪ H1 consists of 2n horizontal strips which are contained in H0 and
H1 . The width of these strips is ≤ 1/5n −→ 0 as n −→ ∞. Thus the intersection of
all of these strips n∈Z+ H n is the product of the interval [0,1] and a Cantor set.
T

The inverse map similarly defines vertical strips, see Figure 5.3. f −1 is really
only defined on the horizontal strips H0 ∪ H1 , but we define

V 0 = V0 ∪V1 = f −1 (H0 ∪ H1 )
5.3. HORSESHOE MAPS AND SYMBOLIC DYNAMICS 85

Figure 5.4: The Horseshoe Map consists of stretching the square in the x direction,
compressing it in the y direction, bending the resulting rectangle into a horseshoe
and intersecting the horseshoe with the original square
86 CHAPTER 5. CHAOTIC DYNAMICS

Figure 5.5: The map of the first two horizontal strips consists of four horizontal
strips.
5.3. HORSESHOE MAPS AND SYMBOLIC DYNAMICS 87

and then define

V n = f −1 (V (n−1) ) ∩ S, n ∈ Z+ ,
so that f −1 stretches in the y direction and compresses in the x analogously to
f , see Figure 5.3. This gives 2n vertical strips of width 1/5n . n∈Z+ V n is now a
T

product of a Cantor set with the y-interval [0, 1] and we define the Smale horseshoe
to be \ \ \
Λ= Hn ∩ Vn = f n (H0 ∪ H1 ).
n∈Z+ n∈Z+ n∈Z

The intersection of the first four vertical and horizontal strips in shown in Figure
5.3.

Lemma 5.3 The Smale horseshoe is invariant with respect to f and f −1 .

Proof: We show that f (Λ) ⊂ Λ, let x ∈ f (Λ) then x ∈ f (H n ) and x ∈ f (V n ), for

all n. Now f (H n ) ∩ f (V n ) ⊂ S, compare Figure 5.3. Thus x ∈ f (H n ) ∩ S = H n+1
and x ∈ f (V n ) ∩ S = V n−1 , for all n. This shows that x ∈ Λ. f −1 is similar. QED

We now establish the relationship of the horseshoe map and symbolic dynam-
ics. Consider the space Σ of binfinite sequences of two symbols {0, 1}. σ ∈ Σ is a
sequence of zeroes and ones infinite in both directions

σ = (· · · , 1, 0, 1, 1, 1, 0, 0, 1, 0, · · · ).

We will denote σ = (σ j ), j ∈ Z, where σ j = 0, 1, and we will be interested in

two operations on Σ. The left shift α : Σ −→ Σ, ασ = (σ j−1 ), j ∈ Z, and the
right shift β : Σ −→ Σ, βσ = (σ j+1 ), j ∈ Z. Now recall the construction of the
horizontal strips in Figure 5.3. We denoted the upper strip H0 and lower strip H1 .
If a point x sits in upper strip we will let σ0 = 0, if it sits in the lower strip we
will let σ0 = 1. Of course this does not give a very precise location of x. The
second iterate gave four horizontal strips. The ones that originated with H0 we
gave the first subscript 0. If strips ended up in H0 they got the second subscript
0, if they ended up in H1 , they got the second subscript 1. Accordingly, σ1 = 0
if x ∈ H0 , σ1 = 1 if x ∈ H1 , and we get the strips, Hσ0 σ1 , σ j = 0, 1, j = 0, 1.
Now the labelling is clear, the nth iterate of H0 ∪ H1 gives Hσ0 ···σn , σ j = 0, 1, and
σ j = 0 indicates that x lies in the strip that ended up in H0 under the jth iterate,
whereas σ j = 1 indicates that x lies in the strip that ended up in H1 under the jth
iterate, see Figure 5.3.
88 CHAPTER 5. CHAOTIC DYNAMICS

Figure 5.6: The map of the first two vertical strips consists of four vertical strips.
5.3. HORSESHOE MAPS AND SYMBOLIC DYNAMICS 89

Figure 5.7: The horseshoe is the intersection of the vertical and horizontal strips.
90 CHAPTER 5. CHAOTIC DYNAMICS

Figure 5.8: The Smale Horseshoe lies with in the 16 squares that are labeled
according to which horizontal and vertical strips they came from.
5.3. HORSESHOE MAPS AND SYMBOLIC DYNAMICS 91

But these strips are nested H 0 ⊃ H 1 ⊃ · · · ⊃ H n and the forward symbol se-
quence
(σ0 , σ1 , σ2 , · · · )
restricts x to lie in a unique line Hσ0 σ1 ··· of the horizontal Cartesian product of
a line with a Cantor set. This is because a nested sequence of strips contains a
unique line. Moreover, we can repeat this labelling with the backward sequence

(· · · , σ−n , · · · , σ−2 , σ−1 )

and the vertical strips Vσ−1 ···σ− j . Then the sequence segment

(σ−( j+1) , · · · , σ−1 , σ0 , · · · , σ j )

restricts x to lie in the rectangle Vσ−1 ···σ−( j+1) ∩ Hσ0 ···σ j and since these rectangles
are also nested and Vσ−1 ··· ∩Hσ0 ··· contains only one point, there is a 1:1 correspon-
dence between the points x ∈ Λ and the biinfinite sequences σ ∈ Σ. This proves
the following lemma.

Lemma 5.4 The map h : Λ −→ Σ, h(x) = σ is 1 : 1 and onto.

We will prove later on that this map is a homeomorphism, but if we take this
to be the case for the time being we get,

Lemma 5.5 The horseshoe map f : Λ −→ Λ is topologically conjugate to shift on

two symbols.

Proof: Consider σ = h(x) and suppose x ∈ Vσ−1 ··· ∩Hσ0 ··· . Then f (x) ∈ Vσ−2 σ−3 ··· ∩
Hσ−1 σ0 σ1 ··· or

h( f (x)) = (· · · σ−2 , σ−1 , σ0 , · · · )

= ασ,

where α is the left shift. This says that

h ◦ f = α ◦ h.

QED

The dynamics on the sequence of two symbols can now be used to give a
complete description of the dynamics on the horseshoe.
92 CHAPTER 5. CHAOTIC DYNAMICS

Fact 5.1 There are infinitely many periodic orbits of all periods in Λ.
Just consider the sequences σ = (σ j ), σ j = 0 and σ = (σ j ), σ j = 1. These
correspond to fixed points of f on Λ, by the topological conjugacy, since ασ = σ
where α is the left shift on Σ. Then consider σ = (σ j ), σ j = 0, if j is even, σ j = 1
if j is odd. This is a periodic orbit of period 1, α2 σ = σ, and there is another one
namely σ j = 1, if j is even, σ j = 0 if j is odd. Thus Λ has exactly two orbits of
period 1 under f . Next we can make a periodic orbit out of a segment of length
three say 011, it is

σ = (· · · , 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, · · · )

and there are 8 = 23 such because we have choice of two symbols for each σ j , j =
0, 1, 2. However, this counts the fixed points (but not the periodic orbits) above so
there are 6 = 8 − 2 periodic orbits of genuine period 2. It is clear that by taking a
segment of length 4 etc, we can construct periodic orbits of any period and count
them all.

Fact 5.2 There exist aperiodic orbits.

This is an easy construction. Take a segment of lenght n add a segment of lenght
2n which is not two copies of the preceeding one then repeat this process in both
forward and backward direction to get an aperiodic orbit.

A Metric
Now we define a metric on the space of sequences
∞
kσ1 − σ2 k = 2
∑ (σ1j − σ2j )2 /2| j| .
j=−∞

Fact 5.3 The periodic orbits are dense in Σ.

Exercise 5.2 Verify Fact 5.3.

Fact 5.4 The left shift has an orbit which is dense in Λ.

We construct the dense orbit in the following manner. First consider all seg-
ments of lenght two. There are four such segments (0, 0), (1, 1), (0, 1) and (1, 0).
We string these segments together and make them the first eights digits σ−8 · · · σ−1
in the negative direction in a sequence σ. σ can have any digits what so ever in
5.3. HORSESHOE MAPS AND SYMBOLIC DYNAMICS 93

the positive direction. Then we take all segments of lenght three and string them
together. They will make the next segment of σ in the negative direction and so
on. Thus we get a sequence that contains arbitrarily large centrally located finite
segments of any sequence σ1 . Now let ε > 0 then there exist integers N and m
such that
kσ1 − αN σk = ∑ (σ1j − σ2j−N )2/2 j ≤ 1/2m−1 < ε,
| j|>m

because the centrally located, i.e. around the zeroth place, segments of σ1 and
αN σ, σ shifted left N times, of lenght 2m are identical.

Lemma 5.6 The map h : Λ −→ Σ, h(x) = σ is a homeomorphism.

The Lemma is proven by the Contraction Mapping Principle, Theorem 2.10,
in Moser [16].

Chaos
Fact 5.5 Λ contains chaotic orbits topologically conjugate to random flips of a
coin.
In other words: Λ contains points the orbits of which under horseshoe map are
topologically conjugate to a shift on a random sequence of two symbols.
We construct a sequence σ by flipping a unbiased coin, i.e. σ j = 0 for heads,
σ j = 1 for tails, j ∈ Z. Λ contains a sequence { f n (x)} n ∈ Z, which is topo-
logically conjugate to σ. This is the precise mathematical meaning of a chaotic
orbit.

Fact 5.6 Λ is uncountable and it contains uncountably many aperiodic orbits.

The sequences σ in Σ are nothing but a binary representation of the real num-
bers and the real numbers are uncountable. The finite sequences that generate the
periodic orbits (of the shifts) in Σ are a binary representation of the rationals that
are countable. Their uncountable complement the irrational numbers are topo-
logically conjugate to aperiodic orbits in Σ.

Example 5.2 Examples of periodic orbits.

a A sequence that terminates
1 1 1
00001110.1 ∼ 1 + + +
22 23 24
94 CHAPTER 5. CHAOTIC DYNAMICS

corresponds to a rational number.

b A sequence that repeats also corresponds to a rational number

1 1 8 1
· · · 100100100.0 ∼ ∑ 23 j − 1 = 1 − 1/8 − 1 = 7 − 1 = 7
j=0

Fact 5.7 The inset and outset of Λ in the square S are the Cartesian product of
the Cantor sets and lines,
\ \
in (Λ) ∩ S = V n , out (Λ) ∩ S = H n.
n∈Z+ n∈Z+ ∪{0}

Proof: Let x ∈ n∈Z+ V n Λ and consider the contraction of the horizontal strips
T

in Figure 5.3. The map f contracts the horizontal strips onto n∈Z+ ∪{0} H n and
T

since each vertical strip contains points in Λ, f (x) ∈ n∈Z+ V n ∩ H 0 ⊂ S. This

T

shows that \ \
y = lim f m (x) ∈ Vn ∩ Hn = Λ
m→∞
n∈Z+ n∈Z+ ∪{0}

or y ∈ Λ. Similarly if

lim f −m (x) ∈ Λ.
\
x∈ H n Λ,
m→∞
n∈Z+∪{0}

QED

The points in in(Λ) can be thought of as semi-infinite sequences σ− =

{· · · , σ−1 } by the map h above, similarly points in out(Λ) can be mapped to semi-
infinite sequences σ+ = {σ0 , σ1 , · · · }. These sequences get closer and closer (in
the norm) to sequences in Σ under the left and right map respectively. Each point
in Λ is the intersection of a unique inset and a unique outset line.

Exercise 5.3

1. The sixteen squares on Figure 5.3 that are the intersections of the four hori-
zontal and the four vertical strips contain each an initial point of a periodic
sequence of period three.

(a) Find the symbolic sequence for each of these periodic orbit and show
in which square it lies.
5.4. THE SMALE-BIRKHOFF HOMOCLINIC THEOREM 95

(b) How many of these are genuine orbits of period three? What are the
others?
(c) Describe how each initial point of a period three orbit moves around
some of the sixteen squares under the horseshoe map.

5.4 The Smale-Birkhoff Homoclinic Theorem

Suppose that the phase space of an ODE contains a hyperbolic stationary solution
and we add small periodic forcing. Then if there is a correct balance of dissipation
and forcing in the system, the associated Poincaré map contains a hyperbolic fixed
point with unstable and stable manifolds that have formed a homoclinic tangle,
see Figure 1.4. The homoclinic tangle was discovered by Poincaré in 1898 and he
used it to prove that the three body problem is not integrable. Roughly speaking
his argument was the familiar argument from complex analysis, that if an analytic
function has the same values on a sequence in the complex plane converging to a
point, then the function must be a constant. Poincaré noticed that the homoclinic
tangle has many sequences of points converging to a point and his argument was
that any integral that was an analytic function of the phase variables and constant
on the sequence had to be trivial. That is to say, the integral had to be the same
constant for all points in the phase space. He also pointed out that the dynamics in
the homoclinic tangle had to be very complex due to the stretching and contracting
close to the hyperbolic point and involved a nonlinear re-injection of the points
into the homoclinic tangle.
During the first half of the nineteen century Birkhoff developed the theory fur-
ther and he was the first one to relate the map to a symbolic sequence. Birkhoff’s
ideas were then taken up by Steven Smale in the 1950’s and 60’s. He formu-
lated and studied the horseshoe map and the associated symbolic dynamics. There
was however no proof that the horseshoe map could exist in the phase space of a
Poincaré map of a flow. This was finally proven in the 1970’s by Charles Con-
ley and Jürgen Moser. They found a mathematically rigorous way of expressing
Poincaré’s observations of a contraction, expansion and nonlinear re-injection in
the vicinity of a hyperbolic point of a Poincaré map. The conclusion of their proof
was that orbits of real flows of ODEs can exhibit the chaotic behaviour formulated
by Birkhoff and Smale.
Conley’s and Moser’s proof is the culmination of one of the major achivements
of 19th century mathematics. It layed the foundation for modern dynamical sys-
tems theory along with the work of Birkhoff and Smale and their mathematical
96 CHAPTER 5. CHAOTIC DYNAMICS

formulation of Poincaré’s ideas reappear in the proofs of most important results

in dissapative dynamical systems during the last three decades.

Theorem 5.3 The Birkhoff-Smale Homoclinic Theorem.

Suppose that a diffeomorphism P : M → M, where M is an n-dimensional man-
ifold, has a hyperbolic fixed point x̄, with a stable W s (x̄) and unstable W u (x̄)
manifold that intersect transversely at some point xo 6= x̄,
W s (x̄) ⊥ W u (x̄), (5.2)
where dimW s + dimW u = n, then M contains a hyperbolic set Λ, invariant under
P, on which P is topologically conjugate to a shift on finitely many symbols.
The transverse intersection is shown in Figure 5.4. Notice that for the Theorem to
make sense at least one of the manifolds W s and W u has to be globally defined.
We will not prove this theorem here but the book by Moser, Stable and Random
Motion in Dynamical Systems [16], contains a very readable proof.

5.5 The Melnikov Method

The hypothesis (5.2) in the Birkhoff-Smale Homoclinic Theorm 5.3 is most eas-
ily proven by use of the Melnikov method. This method consists of computing
a function M(to ) called the Melnikov function which measures the distance be-
tween the stable and unstable manifolds. When M(to ) has simple zeroes then
these manifolds cross transversely. Consider the first order system
ẋ = f (x) + εg(x,t), x ∈ R2 , t ∈ R (5.3)
and assume that the perturbation g is periodic in time, g(x,t + T ) = g(x,t), and
that the unperturbed system
ẋ = f (x)
has a hyperbolic stationary solution x. Also assume that the stable and unstable
manifolds of x form a homoclinic (or a heteroclinic) loop xo (t), see Figure 5.5.
The following lemma is proven in Guckenheimer and Holmes [11].

Lemma 5.7 The stable and unstable manifolds of the Poincaré map of (5.3) in-
tersect transversely if and only if the Melnikov function
Z ∞
M(to ) = f (xo (t − to )) ∧ g(xo (t − to ),t)dt (5.4)
−∞
has simple zeroes.
5.5. THE MELNIKOV METHOD 97

Figure 5.9: The unstable manifold must intersect the stable manifold transversely
in a point x0 6= x̄.
98 CHAPTER 5. CHAOTIC DYNAMICS

Figure 5.10: The homoclinic loop connect the stationary solution to itself W s =
W u.
5.6. TRANSIENT DYNAMICS 99

Exercise 5.4 Show that the Poincaré Map of the damped and driven Duffing’s
Equation
ẍ + δẋ − x + x3 = ε cos(t)
has a Smale Horseshoe in its phase space if
ε 4 cosh(π/2)
>
δ 321/2 π

5.6 Transient Dynamics

The horseshoe map is not a diffeomorphism of the square since f (S) * S, but we
can extend it to a diffeomorphism on the two-sphere S2 . This is done in two steps.
The first step is to extend the square into a soccer stadium so as to include the
whole horseshoe on bottom portion on Figure 2. The results is shown on Figure
5.6. But then we must also extend the map to the portions of the stadium labelled
A and E. This is most conveniently done by placing a sink p in E, outside the
horseshoe, and by letting all the points on the outside perimeter of A be mapped
into A. Then we extend the stadium to a disk so that the map is directed invard on
the perimeter of the disk, see Figure 5.6. This concludes the first step. The second
step is then to make the disk into a spherical cap, to cover the northern hemisphere
and add another cap with a source q at the south pole, see Figure 5.6. It can be
shown that the extended map F, F|S = f is a diffeomorphism of S2 .
Our main interest in the map F is to use it to figure out what happens to all
the points of S, also those that get mapped outside S by f . The following theorem
accounts for all points of S that do not lie on Λ. We let inΛ denote the inset of Λ =
{x ∈ S2 | limn→∞ F(x) ⊂ Λ} and outΛ, the outset of Λ = {x ∈ S2 | limn→∞ F −n (x) ⊂
Λ}. These are not manifold because of the Cantor set structure mentioned above,
however they generalize the notion of stable and unstable manifolds respectively,
see Fact 5.7 in Section 5.3.

Theorem 5.4 All points x of S n∈Z+ V n , where n∈Z+ V n is the vertical Carte-
T T

sian product of a Cantor set and the y-interval [0, 1], eventually approach the fixed
point in E,
lim F n (x) = p.
n→∞

Proof: Consider Figure 5.6. It shows that the regions A, B, D, and E are all
mapped into E in one iteration. All points in E are attracted to p. C is mapped into
A, but all of A gets mapped into E so C gets mapped into E in two iterations. Now
100 CHAPTER 5. CHAOTIC DYNAMICS

Figure 5.11: All points in a neighborhood, except the inset (stable manifold), of
the horseshoe, are eventually mapped to the sink.
5.6. TRANSIENT DYNAMICS 101

consider the two vertical strips V0 and V1 . The strips V 1 = V00 ∪V01 ∪V10 ∪V11 are
mapped onto V 0 = V0 ∪V1 in one iteration. This means that the points in V 0 V 1
must be mapped into B, C or D and therefore into E in at most three iterations, by
the above arguments. Similarly, V 1 V 2 is mapped into E in four iterations of F
etc. The remainder that does not get mapped into E is the intersection of all the
vertical strips n∈Z+ V n .
T
QED
102 CHAPTER 5. CHAOTIC DYNAMICS
Chapter 6

Center Manifolds

Recall from Chapters 1 and 4 the linear stable and unstable manifolds Es and Eu
of a hyperbolic stationary solution. Now we will also consider stationary solution
with pure imaginary eigenvalues and the linear center manifold Ec , spanned by the
eigenvectors of these pure imaginary or zero eigenvalues. The following theorem
give the existence of the nonlinear center manifold W c tangent to Ec .

Theorem 6.1 The Center Manifold Theorem

Consider the systems of equations

ẋ = Ax + f (x, y), x ∈ Rl , y ∈ Rm ,
ẏ = By + g(x, y), l + m = n,

and assume that A has no pure imaginary (or zero) eigenvalues whereas B has
only pure imaginary (or zero) eigenvalues. Moreover, assume that f and g are
Ck (Rn ) functions so that
k f (x, y)k kg(x, y)k
lim = 0 = lim .
k(x,y)k→0 k(x, y)k k(x,y)k→0 k(x, y)k

Then there exists a neighborhood U of the origin in Rn and a center manifold W u

which is tangent to Ec at the origin, or, in other words, there exists a Ck function,

h : Πc (U) −→ Es × Eu ,

Πc (U) being the projection of U onto Ec , whose graph is W c and

hc (0) = 0, Dy hc (0) = 0.

103
104 CHAPTER 6. CENTER MANIFOLDS

Corollary 6.1 The Center manifold x = h(y) is defined by the equation

h0 (y) [By + g(h(y), y)] = Ah(y) + f (h(y), y)

where h0 = Dy h.

Proof: Substitute x = h(y) into the equation for x and use the equation for y to
eliminate ẏ. QED

Example 6.1 Consider the system

ẋ = −x + x2 − y2
ẏ = εy − y3 + xy

Here A = −1 and B = ε so B only has a center manifold at ε = 0. However, if we

write the equations in the form

ẋ = −x + x2 − y2
ẏ = εy − y3 + xy
ε̇ = 0

then there is a two dimensional center manifold. We let x = h(y, ε) and substitute
into the first equation

Dy hẏ + Dε hε̇ = −h + h2 − y2

or
h0 (y) εy − y3 + xy = −h + h2 − y2 .

Now let

h(y, ε) = ay2 + byε + cε2 + O (y, ε)3

Dy h = 2ay + by + O (y, ε)2

2ay + by + O(y, ε)2 εy − y3 + hy = −ay2 − byε − cε2 − y2 + O(y, ε)4

O(y, ε)3 = y2 (a + 1) + bε + cε2 + O(y, ε)3 .

This implies that a = −1 and b = 0 = c, or

h(y, ε) = −y2 + O (y, ε)3 .

 105

The higher order terms are computed by substituting in

h(y, ε) = −y2 + ay3 + by2 ε + cyε2 + dε3 + O(y, ε)4
and solving for the coefficients a, b, c, d. Now the flow on the center manifold is
determined by the equation
ẏ = εy − y3 + hy = εy − 2y3 + O(y, ε)4
ε̇ = 0
where we have substituted h into the center manifold (y) equation.
The example was taken from Carr [6].

Example 6.2
The Lorenz equation
ẋ = σ(y − x)
ẏ = ρx − y − xz
ż = xy − βz
where σ, ρ and β are positive constants, have a stationary solution at the origin
(x, y, z) = (0, 0, 0). It was shown in Section 2.3 that this stationary solution is
stable if ρ ≤ 0 so we let ρ = 1 + µ, then µ is a bifurcation parameter. We will now
compute the center manifold of the Lorenz equation. First write the system in the
form
ẋ = σ(y − x)
ẏ = (1 + µ)x − y − xz
ż = xy − βz
µ̇ = 0
where we have added µ as a variable. Now the equations can be written in the
form
ż = Az + f (z) (6.1)
where A is the matrix
 
−σ σ 0 0
 1 −1 0 0 
A= 
 0 0 −β 0 
0 0 0 0
106 CHAPTER 6. CENTER MANIFOLDS

The eigenvalues of A are 0, −(σ + 1), −β, 0 and the corresponding eigenvectors
are        
1 −σ 0 0
 1   1   0   0 
 ,
 0   0 , 1 , 0 .
    

0 0 0 1
Now let z = Sy where S is the transformation matrix whose columns are the eigen-
vectors, then
      
x 1 −σ 0 0 y1 y1 − σy2
 y   1 −1 0 0   y2   y1 + y2 
 =
  y3  = 
   
 z   0 0 1 0 y3 
µ 0 0 0 1 w w

and the equation (6.1) gets transformed into the equation

 
y1 − σy2
 y1 + y2 
ẏ = By + S−1 f  
 y3 
w

where
 1 σ 
1+σ 1+σ 0 0
−1 1
0 0 
S−1 = 

1+σ 1+σ 
 0 0 1 0 
0 0 0 1
and  
0 0 0 0
 0 −(σ + 1) 0 0 
B = S−1 AS =  
 0 0 −β 0 
0 0 0 0
Now    σ 
y1 − σy2 1+σ (y1 − σy2 )(w − y3 )
1
1+σ (y1 − σy2 )(w − y3 )
 y1 + y2  
S−1 f 

= 
 y3   (y1 − σy2 )(y1 + y2 ) 
w 0
 107

This gives the equation

   σ 
0 0 0 0 1+σ (y1 − σy2 )(w − y3 )
 0 −(σ + 1) 0 0  1
1+σ (y1 − σy2 )(w − y3 )
 
ẏ =  y+ 
 0 0 −β 0   (y1 − σy2 )(y1 + y2 ) 
0 0 0 0 0

This equation shows that the Lorenz equations have a two-dimensional center
manifold at the origin and a two dimensional stable manifold. By the Center
Manifold Theorem there exists a hc ∈ Ck such that

hc : Π(U) → E s

in a neighborhood U of the origin, with hc (0, 0) = (0, 0) and Dhc (0, 0) vanishing.
We let    
y2 h2 (y1 , w)
= h(y1 , w) =
y3 h3 (y1 , w)
The equation determining the center manifold is
 σ
   
D(y1 ,w) h 1+σ (y1 − σh2 )(w − h3 ) =
−(1 + σ) 0 y2
0 0 −β y3
1
 
+ 1+σ (y1 − σh2 )(w − h3 )
(y1 − σh2 )(y1 + h3 )

We approximate h by a power series using that the constant and linear terms must
vanish

h2 = a2 y21 + b2 y1 w + c2 w2 + d2 y31 + e2 y21 w + f2 y1 w2 + g2 w3 + O(y1 , w)4

h3 = a3 y21 + b3 y1 w + c3 w2 + d3 y31 + e3 y21 w + f3 y1 w2 + g3 w3 + O(y1 , w)4

The above equation for the center manifold can be written as two equations

(6.2)
σ ∂h2 1
(y1 − σh2 )(w − h3 ) = −(1 + σ)h2 + (y1 − σh2 )(w − h3 )
1 + σ ∂y1 1+σ
σ ∂h3
(6.3) (y1 − σh2 )(w − h3 ) = −βh3 + (y1 − σh2 )(y1 + h3 )
1 + σ ∂y1
108 CHAPTER 6. CENTER MANIFOLDS

We compute the derivatives of h using the approximation above and substitute

the approximation of h2 and h3 into the equations (6.2) and (6.3). This gives the
equations
σ
(2a2 y21 w + b2 y1 w2 ) + O(y1 , w)4
1+σ
(6.4) = −(1 + σ)(a2 y21 + b2 y1 w + c2 w2 + d2 y31 + e2 y21 w + f2 y1 w2 + g2 w3 )
1
+ (y1 w − a3 y31 − b3 y21 w − c3 y1 w2 − σa2 wy21 − σb2 y1 w2 − σc2 w3 )
1+σ

σ
(2a3 y21 w + b3 y1 w2 ) + O(y1 , w)4
1+σ
(6.5) = −β(a3 y21 + b3 y1 w + c3 w2 + d3 y31 + e3 y21 w + f3 y1 w2 + g3 w3 )
+(y21 w + a3 y31 + b3 y21 w + c3 y1 w2 − σa2 y31 − σb2 y21 w − σc2 y1 w2 )

Equating coefficients of the powers of y1 and w on both sides of these equations

produces the values of the coefficients
1 1 1 1
a2 = 0, a3 = , b2 = , b 3 = , c2 = 0, c3 = 0, d2 = − ,
β (1 + σ)2 β β(1 + σ)2
1 1 σ2 − σβ − 1 2σ
d3 = 2 , e2 = − 2
, e3 = 2
, f2 = − ,
β β(1 + σ) β(1 + σ) (1 + σ)4
σ
f3 = − 3 , g2 = 0, g3 = 0
β (1 + σ)
Substituting these values into the approximation gives the center manifold
1 1 1 2σ
h2 (y1 , w) = 2
y1 w − 2
y31 − 2
y21 w − y1 w2
(1 + σ) β(1 + σ) β(1 + σ) (1 + σ)4
1 1 1 σ2 − σβ − 1 2 σ
h3 (y1 , w) = y21 + y1 w + 2 y31 + y 1 w − y1 w2
β β β β(1 + σ)2 β3 (1 + σ)

up to terms of order (y1 , w)4 . Substituting these expressions into the differential
equation for y1 then gives the flow on the center manifold
σ 1 1 σ
ẏ1 = (y1 w − y31 − y21 w − y1 w2 ) + O(y1 , w)4 (6.6)
(1 + σ) β β (1 + σ)2
We will show below that this equation gives a pitchfork bifurcation as µ = w
increases through zero.
Chapter 7

Bifurcation Theory

Consider the system

ż = Cz + r(z),
where r ∈ Ck and C is a n × n matrix. If C has some eigenvalues on the pure
imaginary axis then we can make a change of coordinates such that the system
can be written

(7.1) ẋ = Ax + f (x, y)
ẏ = By + g(x, y),

f , g ∈ Ck and f , g = O(k(x, y)k2 ), and the eigenvalues of A are pure imaginary,

σ(A) = σc (A), whereas B has no pure imaginary eigenvalues σc (B) = {0}. Here
σ denotes the spectrum and notice that we have switched the x and y from the
statement of the Center Manifold Theorem 6.1. Moreover, the Center Manifold
Theorem says that there exists a center manifold Wloc c , given by y = h(x), tangent

to the linear center subspace Ec at the origin where h ⊂ Ck . We compute h(x) in a

power series in x by recipe

Dx h [A + f (x, h)] = Bh + g(x, h),

given in Corollary 6.1, and then we can observe what bifurcations take place on
c as the coefficients of the h expansion vary. It turns out that if dim x = 1 or 2
Wloc
one can tell the whole story, these are called the codimension 1 and 2 bifurcations
respectively. The codimension 3, dim x = 3, case is much more complicated and
is still unresolved. For higher dimensional cases, dim x > 3, there is not much
that can be said in general unless symmetries are present so that the bifurcations
take place on lower-dimensional subspaces.

109
110 CHAPTER 7. BIFURCATION THEORY

Figure 7.1: The curve of stationary solutions

7.1 Codimension One Bifurcations

We assume that A in (7.1) is a 1 × 1 matrix and that B has no pure imaginary
eigenvalues, then by the Center Manifold Theorem there exists a center manifold
c , given by y = h(x), such that the flow on W c is determined by the equation
Wloc loc

ẋ = f (x, h(x)).

Namely, since A has only one eigenvalue it must be zero and this says that in the
codimension one case the flow is given by a one-dimensional equation. We are
interested in how the flow changes with parameters in the problem and therefore
consider the one dimensional equation

ẋ = f (x, µ) (7.2)

where we have allowed f to depend on the parameter µ as well as x.

The point x = 0, µ = 0 is a stationary solution of (7.2) if and only if

f (0, 0) = 0,
7.1. CODIMENSION ONE BIFURCATIONS 111

and we assume that this is the case by translating, x → x − x0 and µ → µ − µ0 , if

necessary. If
Dx f (0, 0) 6= 0
then by the implicit function theorem there is a C1 branch of stationary solutions
(x, µ) through the origin, see Figure 7.1, such that
f (x, µ) = 0.
A necessary condition for a bifurcation to take place is therefore
Dx f (0, 0) = 0 (7.3)
and we will now assume this is the case and that the two degeneracy conditions
Dµ f (0, 0) 6= 0, D2x f (0, 0) 6= 0, (7.4)
also hold.

7.1.1 The Saddle-Node Bifurcation

The two conditions (7.3) and (7.4) give the generic codimension one bifurcation.
It is called the saddle-node bifurcation, see Figure 7.1.1 and Figure 7.1.1 and, and
is the bifurcations that we should expect to see in the codimension one case unless
there are some extra symmetries present. First we notice that the vector field will
vanish
f (x, µ) = 0
on a branch x(µ) of stationary solutions. A differentiation of f (x, µ) = 0 with
respect to x gives
dµ
Dx f (0, 0) + Dµ f (0, 0) = 0,
dx
so
dµ
= 0,
dx
at the bifurcation point. A second differentiation gives
 2
2 dµ 2 dµ d2µ
Dx f (0, 0) + 2Dµx f (0, 0) + Dµ f (0, 0) + Dµ f 2 = 0,
dx dx dx
or
d2µ D2x f (0, 0)
=− .
dx2 Dµ f (0, 0)
112 CHAPTER 7. BIFURCATION THEORY

X X

µ µ

Figure 7.2: The supercritical saddle-node bifurcation, α = β = ±1.

This says that locally the bifurcation curve is a parabola centered on the µ-axis.
Moreover, since the stability of the bifurcating solutions (x, µ) are determined by
the linearized equation (7.2),

ẏ = Dx f (x, µ)y (7.5)

and
dµ
Dx f (x, µ) = −Dµ f (x, µ)
dx
we get the stability information in Table 7.1.1. Now there are 4 possible bifurca-
tion diagrams which are illustrated on Figures 7.1.1 and 7.1.1.
Table III.1

Dµ f > 0 Dµ f < 0
µ0 >0 Stable Unstable
µ0 <0 Unstable Stable

These bifurcations take place for the function

f (x, µ) = αµ − βx2 , α, β = ±1 (7.6)

and the supercritical cases on Figure 7.1.1 correspond to α = β = 1 and α =

β = −1 respectively, whereas the subcritical cases on Figure 7.1.1 correspond to
7.1. CODIMENSION ONE BIFURCATIONS 113

X X

µ µ

Figure 7.3: The subcritical saddle-node bifurcation, α = ±1, β = ∓1.

α = 1, β = −1 and α = −1, β = 1, respectively. In fact (7.6) is called the normal

form of the saddle-node bifurcation because all codimension one flows exhibiting
this bifurcation can be reduced to the form (7.6).

7.1.2 A Transcritical Bifurcation

Now suppose for a moment that f (0, µ) = 0 so that x = 0 is a stationary solution
of (7.2). This means that
Dµ f (0, 0) = 0 (7.7)
and now we impose this condition along with the nondegeneracy condition
D2x f (0, 0) 6= 0. (7.8)
We have to impose another condition in addition to (7.8), to get branches of
stationary solutions. This condition says in the transcritical case that the origin
(0, 0) is a saddle-point of f (x, µ). Notice that (7.3) and (7.8)
Dx f (0, 0) = 0, Dµ f (0, 0) = 0
state that (0, 0) is a critical point of f but this critical point cannot be a maximum,
if we are to have branches of stationary solutions. We expand f (x, µ) in a Taylor
series about the origin to make this explicit
f (x, µ) = D2x f (0, 0)x2 + 2Dxµ f (0, 0)xµ + D2µ f (0, 0)µ2 + O(x, µ)3 .
114 CHAPTER 7. BIFURCATION THEORY

X X

µ µ

Figure 7.4: The transcritical bifurcations.

The origin is a saddle if and only if

2
D = Dxµ f (0, 0) − D2x f (0, 0)D2µ f (0, 0) > 0, (7.9)
and then we get two branches of solutions of f (x, µ) = 0, namely
 √  µ
x± = −Dxµ f ± D
D2x f
where D is the discriminant (7.9). Notice that in general x = 0 is not a branch of
stationary solutions but it becomes one if
D2µ f (0, 0) = 0. (7.10)
However, the conditions (7.8) and (7.9) imply that there are two branches of sta-
tionary solutions going through the origin. The stability of those two branches is
given by the linearization of the equation (7.2),
ẏ = 2 D2x f x + Dxµ f µ y + O(x, µ)2


√
= ±2 Dµy + O(x, µ)2 .
Thus the x+ and x− branch have opposite stability and since µ changes its stability
at the origin, the two branches exchange stability at the origin.
The transcritical bifurcation is exhibited by the vector field
f (x, µ) = x(αµ + βx), α, β = ±1
The two cases on Figure 7.1.2 correspond to α = −1 and α = +1 respectively.
We get two bifurcation diagrams see Figure 7.1.2.
7.1. CODIMENSION ONE BIFURCATIONS 115

x x

µ µ

x x

µ µ

Figure 7.5: The pitchfork bifurcations.

7.1.3 A Pitchfork Bifurcation

We now let
D2x f (0, 0) = 0, (7.11)
then the Taylor expansion of f around the origin becomes

f (x, µ) = 2Dxµ f (0, 0)xµ + D3x f (0, 0)x3

+3D2x Dµ f (0, 0)x2 µ + O(µ2 ).

This means that for µ small we can get 3 branches of stationary solutions. Namely,

f (x, µ) = x D3x f x2 + 3D2x Dµ f xµ + 2Dxµ f µ + O(µ2 )

and we get the 3 solutions

s
−2Dxµ f (0, 0)µ
x0 = 0 and x± = ± . (7.12)
D3x f (0, 0)
116 CHAPTER 7. BIFURCATION THEORY

This makes the second non-degeneracy condition

Dxµ f (0, 0) 6= 0 (7.13)

clear. A differentiation of f (x, µ)/x similar to the one performed in the saddle-
node case above also shows that
dµ d2µ D3x f
= 0 and = − 6= 0,
dx dx2 3Dx Dµ f

for the x± branches. This says that these branches form a parabola along the µ
axis in the µ − x plane, see Figure 7.1.3.
The stability is determined by the linearization of (7.2) about these branches.
For the x = 0 branch we get

ẏ = 2(Dxµ f (0, 0)µ)y

and for the x± branches we get

ẏ = −4(Dxµ f (0, 0)µ)y.

Thus the stability of the x = 0 is opposite to that of the x± branches and determined
by the signature of Dxµ f (0, 0), see Figure (7.1.3). The signatures of D3x f and Dxµ f
determine in which µ half-plane we get three branches of stationary solutions,

sign D3x f (0, 0) 6= sign Dxµ f (0, 0)

is called the supercritical and

sign D3x f (0, 0) = sign Dxµ f (0, 0)

is called the subcritical case, see Figure 7.1.3. In the former case we get three
branches for µ positive in the latter case we get three branches for µ negative.
Now the canonical example or normal form exhibiting pitchfork bifurcations
is
f (x, µ) = αµx − βx3 , α, β = ±1.
The cases on Figure 7.1.3 correspond to α = 1 = β, α = −1 = β, α = −1, β =
+1 and α = +1, β = −1, respectively.
7.1. CODIMENSION ONE BIFURCATIONS 117

Example 7.1

Recall the equation (6.6) describing the flow on the center manifold of the
Lorenz equations

σ 1 1 σ
ẏ = (yµ − y3 − y2 µ − yµ2 ) + O(y, µ)4 = f (y, µ) (7.14)
(1 + σ) β β (1 + σ)2

We compute the derivatives give us the criteria for a bifurcation and the type of
bifurcation of the stationary solution at the origin (y, µ) = (0, 0).

∂f σ 3 2 σ
= (µ − y2 − yµ − µ2 ) = 0, at (0, 0)
∂y 1 + σ β β (1 + σ)2
∂f σ 1 2σ
= (y − y2 − yµ) = 0, at (0, 0)
∂µ 1 + σ β (1 + σ)2
∂2 f σ 6 2
2
= (− y − µ) = 0, at (0, 0)
∂y 1+σ β β

The first line shows that there is a bifurcation point at the origin. The second line
shows that it is not a saddle-node bifurcation. The third line shows that it is not a
transcritical bifurcation. Next we show that the non-degeneracy conditions for a
pitchfork bifurcation at the origin are satisfied.

∂2 f σ 2 2σ σ
= (1 − y − 2
µ) = 6= 0,
∂y∂µ 1+σ β (1 + σ) 1+σ
∂3 f σ 6
3
= − 6= 0
∂y 1+σ β

Thus the Lorenz equation have a pitchfork bifurcation at the origin and now we
show that it is supercritcal, namely

∂2 f ∂3 f
sign 6= sign 3
∂y∂µ ∂y

More information on the bifurcation theory of ODEs and proofs can be found
in Iooss and Joseph [13] and for PDEs in Chow and Hale [7]. The reduction to
normal forms is performed in Arrowsmith and Place [3].
118 CHAPTER 7. BIFURCATION THEORY

Figure 7.6: The Poincaré Map.

7.2 The Poincaré Map

We discussed in Definition 2.7 how one can produce a map from a flow by taking
the time-T map. It is important to realize that not all such maps are equivalent or
topologically conjugate see Arrowsmith and Place [3]. For example if the flow
has a periodic orbit a of period T then the time-T map has a fixed point but the
time- T2 has a periodic orbit of period one. Consequently their topological type is
different.
Poincaré constructed a map by taking a hyperplane transverse to a periodic
orbit and defining the map to be the return map to the hyperplane. These maps are
parametrized by the intersection point of the hyperplane and the periodic orbit and
they are all topologically conjugate for a sufficiently small neighborhood of the
periodic orbit, see Figure 7.2. Now we generalize Poincaré’s construction slightly,

Definition 7.1 Let {{T (t)}, X} be a flow, then we say that a codimension one
hypersurface Γ ⊂ X is a transversal, in a neighborhood U of u ∈ Γ, if every orbit
u(t) = T (t)uo meets Γ again and the vector field u̇, at u, is not tangent to Γ. Γ is
7.3. THE PERIOD DOUBLING BIFURCATION 119

a global transversal if T (t)uo meets Γ for arbitrarily positive or negative t.

Not every flow has a transversal but if it does we define the first return map in
the following way:

Definition 7.2 The Poincaré map of Γ ∩U is defined to be

P(w) = T (τ)w, w ∈ Γ ∩U, (7.15)

where τ(w) is the first time that the orbit u(t) = T (t)w returns to Γ ∩U, U being
a neighborhood of w in X.

7.3 The Period Doubling Bifurcation

Now consider a map

xm+1 = f (xm , µ), x, µ ∈ R, m ∈ Z, (7.16)

and suppose
f (0, 0) = 0 and Dx f (0, 0) = −1.
This gives rise to a bifurcation which is not possible for one-dimensional flows,
namely a branch of fixed points bifurcates into a periodic orbit. The names comes
from the fact that if the map is the Poincaré map of a flow and the fixed point of the
map corresponds to the periodic orbit of the flow, of period one, then the periodic
orbit of the map corresponds to a periodic orbit of the flow with period two. Now
the implicit function theorem implies that there is a branch of fixed points going
through the origin. We analyze the bifurcation by considering the second iteration
of the map
xm+1 = f 2 (xm , µ)
and consider the flow
ẋ = g(x, µ) = f 2 (x, µ) − x.
The derivatives at the origin are

Dx g = (Dx f )2 (0, 0) − 1 = 0

by chain rule,

Dµ g = Dx f Dµ f (0, 0) = 0,
120 CHAPTER 7. BIFURCATION THEORY

if Dx f = −1, and Dµ f = 0.
In addition
D2x g = D2x f Dx f (Dx f + 1) = 0,
but
D2µx g = 2Dx f Dxµ f (0, 0) + D2x f Dµ f = −2Dxµ f (0, 0) 6= 0,
and
2
D3x g = Dx f D3x f 1 + (Dx f (0, 0))2 + 3 D2x f Dx f (0, 0)
  
n 2 o
= − 2D3x f (0, 0) + 3 D2x f (0, 0)

6= 0

in general. Thus by the analysis of the pitchfork bifurcation in the previous sec-
tion,
xm+1 = f 2 (xm , µ)
has a pitchfork bifurcation at the origin. This means that if we denote by x+ the
top and x− the bottom pitchfork branch, then

x+ = f (x− , µ)

and
x− = f (x+ , µ).
The reasoning is that the second iterate must be a stable fixed point of f 2 , but there
are only two such fixed points. Moreover,

x− = f (x− , µ)

is impossible because then f would have another branch of fixed points going
through the origin. The following theorem holds,

Theorem 7.1 Suppose that f ∈ C3

Dxµ f (0, 0) 6= 0

and 2
2D3x f (0, 0) + 3 D2x f (0, 0) 6= 0,


then the map (7.16) has a period-doubling (also called a flip) bifurcation at the
origin. There exists a stable branch of fixed points for µ < 0 that becomes unstable
for µ > 0, at the origin, and there exists a branch of stable periodic orbits for µ > 0.
7.4. THE HOPF BIFURCATION 121

7.4 The Hopf Bifurcation

The Hopf bifurcation is the generic codimension two bifurcation from a stationary
solutions. The stationary solution becomes unstable and bifurcates to a periodic
orbit, see Figures 7.4 and 7.4.
Theorem 7.2 Suppose that the system
ż = F(z, µ), z ∈ R2 (7.17)
has two complex (conjugate) eigenvalues λ = α ± iβ crossing the pure imaginary
axis at µ = 0, with positive speed,
Reλ
Reλ(0) = 0, > 0, (7.18)
dµ
Then if the origin is a stable stationary solution, (this is the case if the coefficient a
below is negative) µ = 0 is a bifurcation point and the bifurcation is supercritical,
x = 0 is a stable stationary solution for µ < 0, and x = 0 is an unstable stationary
solution, encircled by a stable periodic orbit, for µ > 0.

Using the theory of Poincaré-Birkhoff normal forms, see for example Arrowsmith
and Place [3], the equation (7.17) can be reduced to an equation in polar coordi-
nates,
ṙ = µr + ar3 + O(r5 , µ)
(7.19) θ̇ = ω + br2 + O(r4 , µ).
The bifurcation is controlled by the r equation and it is the pitchfork equation that
we analyzed above. The quantity a is computed from the vector field F = ( f , g),
1
a = ( fxxx + fxyy + gxxy + gyyy )
16
1
(7.20) + ( fxy ( fxx + fyy ) − gxy (gxx + gyy ) − fxx gxx + fyy gyy ) ,
16
see for example Guckenheimer and Holmes [11]. If a < 0 then the bifurcation
is supercritical and we get a branch of stable stationary solution for µ < 0, that
becomes unstable at the origin and throws off a stable periodic orbit. These orbits
form a paraboloid whose center is the branch of unstable stationary solutions for
µ > 0, see Figure 7.4. For µ > 0, the origin in x-space is encircled by a unique
stable periodic orbit whose size and period changes continuously with µ, see [3].
If a > 0 the bifurcation is subcritical and the stability reverses, the unstable
periodic orbit encircles the origin, in x-space, see Figure 7.4, for µ < 0.
122 CHAPTER 7. BIFURCATION THEORY

Figure 7.7: The Supercritical Hopf Bifurcation.

Figure 7.8: The Subcritical Hopf Bifurcation.

Chapter 8

The Period Doubling Cascade

8.1 The Quadradic Map

We consider the one-dimensional quadratic map
xm+1 = f (xm , µ), x, µ ∈ R, m ∈ Z,
where
f (x, µ) = 1 − µx2 .
We will use this map to illustrate the phenomena that occur for a range of µ-values.
These are in fact typical for a whole class of one-dimensional maps and this will
be made precise below. The iterations are illustrated in Figure 8.1 . We consider
the map in the interval x ∈ [−1, 1] and f forms a hump in this interval. The fixed
points x of f are the points where the line y = x intersects the graph y = f (x).
These fixed points are stable if | f 0 (x)| < 1, unstable if | f 0 (x)| > 1, i.e. if the slope
of f at the fixed point is less or greater than 45◦ respectively.
It turns out that there is a natural condition that the function f has a negative
Schwartzian derivative which implies that if the map has a stable fixed point then
it is unique. The Schwartzian derivative is
 00 2
f 000 (s) 3 f (s)
SD( f (x)) = 0 −
f (x) 2 f 0 (x)
see [8], and the condition is that SD( f (x)) < 0. Moreover, then the set of points
which is not attracted to the fixed point has Lebesgue measure zero. But we may
ask if there are maps which have no stable fixed points, see Figure 8.1, and what
happens then to the iterations of most points.

123
124 CHAPTER 8. THE PERIOD DOUBLING CASCADE

Figure 8.1: The Stable Quadradic Map.

8.1. THE QUADRADIC MAP 125

Figure 8.2: The Unstable Quadradic Map.

126 CHAPTER 8. THE PERIOD DOUBLING CASCADE

If we consider the bifurcation diagram of x, see Figure 8.1 as a function of µ

then there is an initial interval where the attractor is a stable fixed point. Then a
period doubling bifurcation occurs and we get an interval where the attractor is a
stable periodic orbit of period one. Then another period doubling bifurcation takes
place and we get a much shorter interval where the attractor is a stable periodic
orbit of period two. This continues and we get a sequence of intervals where the
attractors are stable periodic orbits of period 4, 8, 16 etc. These bifurcation points
{µn } converge at µ∞ ,
lim µn = µ∞ ,
n→∞
see Figure 8.1, where the attractor becomes a singularly supported strange attrac-
tor, see Figure 8.1. Notice that in each interval the periodic orbit from the previous
interval survives but becomes unstable. Thus each interval contains unstable peri-
odic orbits of all the previous periods. Moreover, at the center of each interval lies
a point where periodic orbit contains zero and the derivative Dx f (0) = 0. These
orbits are called the superstable periodic orbits. What can we say about the region
beyond µ∞ ? It is known that the set of µ0∞ for which there exists no stable pe-
riodic orbit has positive Lebesgue measure and the slightly smaller set for which
there is sensitive dependance on initial conditions also has positive Lebesgue mea-
sure. It has been proven more recently that the still smaller set where there exists
absolutely continuous invariant measures, see Figure 8.1, has positive Lebesgue
measure. If a map has a density that is an absolutely continuous function ρ with
respect to Lebesgue measure and the measure defined by this density

dµ = ρ(x)dx

is invariant with respect to the map

dµ( f n (S)) = dµ(S)

where S ⊂ [−1, 1] is any subset of the interval. Then we say that the map possesses
an absolutely continuous invariant measures. We shall think about such maps as
being strongly chaotic because they also possess positive Lyapunov exponents and
the associated sensitive dependance on initial conditions.
Now we describe how to construct a map with sensitive dependance on initial
conditions. The method is to let f fall on an unstable fixed point after a few it-
erations of the map. The reason is that f 0 (0) = 0 and this is the greatest stability
one can achieve in any neighborhood. If that neighborhood falls on the neigh-
borhood of an unstable fixed point after a few iterations, then the points in the
8.1. THE QUADRADIC MAP 127

Figure 8.3: The Period Doubling Cascade.

128 CHAPTER 8. THE PERIOD DOUBLING CASCADE
8.1. THE QUADRADIC MAP 129

Figure 8.5: The histogram and support of the strange attractor with an absolutely
continuous invariant measure at µ = 2.
130 CHAPTER 8. THE PERIOD DOUBLING CASCADE

neighborhood must disperse. We illustrate this in Figure 8.1. where

f (x) = 1 − 1.544x2 .

The origin is mapped onto the unstable fixed point x in three iterations and nearby
points must disperse. In Figure 8.1 we show the corresponding histogram and the
support of the strange attractor.

8.2 Scaling Behaviour

We now consider a numerical observation made by Feigenbaum. Let 2n denote
the period and µn the µ-value where a bifurcation from the period 2n−1 to 2n
−µn−1
occurs. Then we list the µn s, the difference µn − µn−1 and the ratios µµnn+1 −µn . The
last column of the table shows a striking feature.

Table III.1
µn −µn−1
n µn µn − µn−1 µn+1 −µn

0 .75 .5
1 1.25 .1180989394 4.233738275
2 1.3680989394 .0259472172 4.551506949
3 1.3940461566 .0055850823 4.645807493
4 1.3996312389 .0011975035 4.663938185
5 1.4008287424 .0002565289 4.668103672
6 1.4010852713 .000054943399 4.668966942
7 1.401140214699 .000011767330 4.669147462
8 1.401151982029 .000002520208 4.669190003
9 1.401154502237 .000000539752 4.669196223
20 1.401155041989

The first column shows that the µn values actually converge to a terminal point
µ∞ . The second column shows that the windows of periodicity get smaller and
smaller. The last column shows that the ratio converge to a constant

δ = 4.66920 · · · .
8.2. SCALING BEHAVIOUR 131

Figure 8.6: The Unstable Quadradic Map.

132 CHAPTER 8. THE PERIOD DOUBLING CASCADE

Figure 8.7: The histogram and support of the strange attractor with an absolutely
continuous invariant measure, covering part of the interval, at µ = 1.544.
8.2. SCALING BEHAVIOUR 133

The bifurcation points µn are given by the scaling formula

(µ0 − µ∞ )
µn = + µ∞
δn
The surprising fact that was discovered by Feigenbaum and proven by Collet-
Eckmann-Lanford is that the constant δ does not depend on the particular map but
is a universal constant for sufficiently smooth map of the interval to itself. In fact
there are two universal constants associated to smooth maps from the interval to
itself, the other being
λ = 0.3995 · · ·
and some geometrical features of the bifurcation diagram in Figure 8.1 scale like
λn .

Definition 8.1 The periodic points µ0n of period 2n , whose periodic orbits contain
zero, are called superstable.
The name superstable comes from the fact that
n
D2 f (0) = 0,

which gives the strongest linear contraction possible for the period 2n . The super-
stable points satisfy the scaling relationship
(µ00 − µ∞ )
µ0n = + µ∞
δn
We plot in Figure 8.2. the period doubling cascade with − log (µ∞ − µ) on the
vertical axis instead of µ. The result is periodic windows of period 2n . An even
more dramatic result is achieved if we scale the x-axis as well. Figure 8.2. shows
the period doubling sequence with the vertical axis x · |µ∞ − µ|−0.59367 , where the
exponent α = 0.59367 is given by the relationship

λ = δα

between the two universal constants. This gives a periodic bifurcation diagram
with a reflection (about the unstable pitchfork branch) symmetry.
Now we consider the other side of µ∞ . The point marked µµ0 is the point where
f 3 (0) is an unstable fixed point and we get a sequence of values µµn such that at
(µµ0 − µ∞ )
µµn = + µ∞
δn
134 CHAPTER 8. THE PERIOD DOUBLING CASCADE

Figure 8.8: The rescaled Period Doubling Cascade.

8.2. SCALING BEHAVIOUR 135

Figure 8.9: The rescaled Period Doubling Cascade.

136 CHAPTER 8. THE PERIOD DOUBLING CASCADE

n
f 3·2 (0) falls on an unstable periodic orbit of period 2n . In the period 3 window f
(3)
will have a periodic orbit of period 3 · 2n−1 at µn , and

(µ3 − µ0∞ )
µ3·2n = n
+ µ0∞
δ

for a new accumulation point µ0∞ . In other words, the scaling holds on both sides
of µ∞ and applies to period doubling of period two on one side and period three
on the other side. Furthermore, the same analysis applies to period tripling,
quadrapling etc. sequences, with new universal constants λ(n) and δ(n), 3, 4, · · · .

Figure 8.2 illustrates the stable and unstable manifold of the renormalization
map in the following theorem.

Theorem 8.1
1
F g(x) = − g ◦ g(−ax),
a
where g(0) = 1 and a is a constant, denotes the period halving map acting on the
space Gε = {g(x) : [−1, 1] → R | g(x) = f (|x|1+ε ), ε ≤ 1}, where f is a bounded
analytic function on [0, 1] satisfying f (0) = 1, ddyf < 0 on [0, 1] and f (1) > −1;
then for ε ≤ 1, F has a unique fixed point gε in Gε , with a negative Schwartzian
derivative. gε is hyperbolic and

dimW u (gε ) = 1 = codimW s (gε ),

where both W u and W s are smooth manifolds.

Corollary 8.1 For each a ∈ [−1, 1] there exists a unique point ga ∈ W u ⊂ Gε such
that ga (1) = −a. W u intersects the hypersurfaces

Σ1 = {g ∈ Gε | g(1) = 0}

and
e1 = {g ∈ Gε | g3 (1) = −g(1)}
Σ
transversely.

For a proof of the Theorem and the Corollary, see Collet and Eckmann [8].
8.2. SCALING BEHAVIOUR 137

Figure 8.10: The stable and unstable manifolds of the period halving map.
138 CHAPTER 8. THE PERIOD DOUBLING CASCADE

Corollary 8.2 Let gµ ⊂ Gε be a continously differentiable family parametrized

by µ, and suppose that gµ intersects W s tranversally with non-zero velocity, at µ∞ .
Then there exists sequence {µ j } and {e
µ j } converging to µ∞ from opposite sides of
W s such that the finite limits

lim δ j (µ∞ − µ j ) and lim δ j (µ∞ − e

µ j)
j→∞ j→∞

exists. The corresponding maps gµ j have a super-stable orbit of period 2 j , and

geµ j admits an absolutely continous invariant measure for j large enough.
 
µ −µ
Remark 8.1 The ratio of the limits lim j→∞ µ∞∞− eµjj is universal, i.e. does not
depend on the particular family gµ in question and we can for example choose the
representative maps
gµ (x) = 1 − µ|x|1+ε .

8.2.1 The Singularly Supported Strange Attractor

Theorem 8.1 and its Corollaries describe what happens on each side of the stable
manifold W s ⊂ Gε , see Figure 8.2. Now we will describe what happens on the
stable manifold W s .

Definition 8.2 Let g ∈ Gε , then J = ∩ J j is a g-invariant Cantor set if

1. The J j s form a nested sequence of closed subsets of [−1, 1]

J0 ⊃ J1 ⊃ · · · ,

0 ∈ J j , for all j, and g(J j ) ⊂ J j .

2. J j+1 consists of 2 j+1 closed intervals formed by deleting an open subinter-

val from all the closed intervals forming J j .

3. g maps the subintervals of J j onto one another, J j ⊂ g(J j ), such that the
action of g on J j is a cyclic permutation of order 2 j .

4. The map g : J → J is an isomorphism and every orbit of g in J is dense in J.

We call J a strange g-invariant Cantor set if in addition, g has a negative

Schwartzian derivative and it satisfies two more conditions:
8.2. SCALING BEHAVIOUR 139

5. For each j = 1, 2, · · · , g has a unique repelling periodic orbit of period 2 j

which does not belong to J j .

6. Each orbit of g either,

(a) lands after finitely many iterates on one of the periodic orbits in 5,
(there are only countably many such orbits), or
(b) converges to J, such that for each j it is eventually contained in J j .

Theorem 8.2 Let g ∈ W u , then g possesses a g-invariant Cantor set; if g ∈ W u has

a negative Schwartzian derivative, then it possesses a strange g-invariant Cantor
set. Moreover, if ν is a probability measure with support(ν)=J, such that each
of the 2 j subintervals of J j are assigned equal weight under ν, then ν is the only
probability measure on J, invariant under g. The dynamical system (ν, g) is er-
godic but not weakly mixing and consequently,

1 n−1
Z
lim ∑ φ(gn (x)) = φ(x)dν,
n→∞ n J
j=0

where x is any point in [−1, 1] whose orbits converge to J and φ is any continuous
function on [−1, 1].
For a proof of the Theorem see Collet and Eckmann [8].
We now define the tools we need for sensitive dependence on initial condi-
tions.

Definition 8.3 A Lyapunov exponent of a one-dimensional map xm+1 =

f (xm , µ), m ∈ N, is the limit
log |Dx f (x)|
λ(x) = lim (8.1)
n→∞ n
Notice that the Lyapunov exponent is defnined for any x in the phase space of the
map not only for fixed points. At a fixed point the Lyapunov exponent equals the
eigenvalue from Lemma 5.1.
A measure m is invariant under a map f if

m(I) = m( f −1 (I))

where I is a measurable set.

The following theorem was proven by Benedickts and Carleson [5].
140 CHAPTER 8. THE PERIOD DOUBLING CASCADE

Theorem 8.3 Let c < log 2 be given. Then there exists a set E of positive
Lebesgue measure close to µ = 2, for the map f (x, µ) = 1 − µx2 , such that the
Lyapunov exponent at x = 1 for µ ∈ E is greater than c.
Corollary 8.3 There exists a small positive number α such that
√
| f j (0)| ≥ e−α j
, for all j ≥ 0 (8.2)
The Corollary shows that the contraction rate of the map at the origin is bounded
from below. This is exactly what one needs to prove Jacobson’s Theorem [14]:
Theorem 8.4 If µ ∈ E so that (8.2) holds, then the map x → f (x, µ) has an in-
variant Borel probability measure that is absolutely continuous with respect to
Lebesgue measure.
The period doubling cascade has applications in all branches of science and
engineering. For applications to superconductors see [4], to electrons in quantum
wells driven by lasers see [10] and [1] and to earthquakes see [9].
Example 8.1 The map x → 1 − 2x2 on [−1, 1] is mapped to the map y → 1 − 2|y|
on [−1, 1] by the homeomorphism
r
4 −1 x+1
y = sin ( )−1 (8.3)
π 2
The Lyapunov exponent of the latter map is obviously λ(y) = 2 because Dy f = ±2
depending on whether y is positive or negative. The latter map has the invariant
measure
1 1
dm = dy (8.4)
π (1 − y2 )1/2
This measure is obviously absolutely continuous with respect to Lebesgue mea-
sure, since
1 1
>0
π (1 − y2 )1/2
and m is a probability measure because
Z 1
1 1 1
2 1/2
dy = sin−1 (y)|1−1 = 1
π −1 (1 − y ) π
Exercise 8.1
1. Show that the map (8.3) maps x → 1 − µx2 to y → 1 − 2|y|.
2. Show that the measure (8.4) is invariant under the map y → 1 − 2|y|.
Appendix A

The Homoclinic Orbits of the

Pendulum

We start with the energy of the pendulum

y2
E= + 1 − cos (x)
2
where y = ẋ and set it equal to the energy of the stationary solution (x, y) = (π, 0),

E(π, 0) = 1 − cos (0) = 2,

y2
+ 1 − cos (x) = 2,
2
y2 x
= 1 + cos (x) = 2 cos2 ,
2 2
by the identity
1
cos2 (z) =
(1 + cos (2z))
2
from trigonometry. This gives the equation
ẋ2 x
= 2 cos2
2 2
or x
ẋ = ±2 cos .
2
ẋ
We let z = 2 to get
ż = ± cos (z).

141
142 APPENDIX A. THE HOMOCLINIC ORBITS OF THE PENDULUM

This first order ODE is separable

dz
= dt
cos (z)
or Z
sec (z)dz = ±(t + t0 ).

This integral is evaluated in calculus

Z
sec (z)dz = ln | sec (z) + tan (z)|

or
ln | sec (z) + tan (z)| = ±(t + t0 )
so exponentiating both sides we get that

sec (z) + tan (z) = e±(t+t0 ) .

This is now an equation that we need to solve for z, to do that we use the trigonom-
etry identity
1 + tan2 (z) = sec2 (z)
or p
sec (z) = 1 + tan2 z.
If we set e±(t+t0 ) = q the equation becomes
q
1 + tan2 (z) + tan (z) = q.

We move tan (z) to the right hand side and square both sides
q
1 + tan2 (z) = q − tan (z)

so
1 + tan2 (z) = q2 − 2q tan (z) + tan2 (z).
Thus
1 = q2 − 2q tan (z)
and solving for tan (z) gives
1
tan (z) = (q − q−1 ).
2
 143

However,
1 e±(t+t0 ) − e∓(t+t0 )
(q − q−1 ) = = ± sinh (t + t0 ).
2 2
Thus
tan (z) = ± sinh (t + t0 )
and
x = 2z = ±2 tan−1 (sinh (t + t0 ))
since both sinh and tan are odd functions. Moreover,
cosh (t + t0 )
ẋ = ±2
1 + sinh2 (t + t0 )
1
= ±2 = ±2 sech(t + t0 )
cosh (t + t0 )

using the trigonometry identity

1 + sinh2 (t + t0 ) = cosh2 (t + t0 ).

We have now shown that the homoclinic orbits are

(x, y) = ±2(tan−1 (sinh (t + t0 )), sech(t + t0 )),

as t −→ ±∞ , we get
(x, y) −→ (±π, 0).
144 APPENDIX A. THE HOMOCLINIC ORBITS OF THE PENDULUM
Bibliography

[1] Batista A, B. Birnir, P. I. Tamborenea, and D. Citrin. Period-doubling and

Hopf bifurcations in far-infrared driven quantum well intersubband transi-
tions. Phys. Rev. B., 68:035307, 2003.

[2] V. I. Arnold. Ordinary Differential Equations. MIT Press, Cambridge, MA,

1991. (Translated and Edited by R. A. Silverman).

[3] D.K. Arrowsmith and C. M. Place. An Introduction to Dynamical Systems.

Cambridge University Press, Cambridge, UK, 1994.

[4] B.Birnir and R.Grauer. The global attractor of the damped and driven sine-
Gordon equation. Comm. Math. Phys., 162:539–590, 1994.

[5] M. Benedicks and L. Carleson. On iterations of 1 − ax2 on (−1, 1). Annals

of Math., 122:1–25, 1985.

[6] J. Carr. Applications of Centre Manifold Theory. Springer, New York, 1981.

[7] S.-N. Chow and J. Hale. Methods of Bifurcation Theory. Springer, New
York, 1982.

[8] P. Collet and J.-P. Eckmann. Iterated Maps on the Interval as Dynamical
Systems. Birkhäuser, Boston, 1980. PPhI Progress in Physics.

[9] B. Erickson, B. Birnir, and I. Lavallée. A model for aperiodicity in

earthquakes. Nonlin. Processes Geophys, 15:1–12, 2008. Available at
http://repositories.cdlib.org/cnls/.

[10] B. Galdrikian and B.Birnir. Period doubling and strange attractors in quan-
tum wells. Phys. Rev. Lett.., 76(18):3308–11, 1996.

145
146 BIBLIOGRAPHY

[11] J. Guckenheimer and P. Holmes. Nonlinear Oscillations, Dynamical Sys-

tems, and Bifurcations of Vector Fields. Springer, 1983.

[12] P. Hartman. Ordinary Differential Equations. Wiley, New York, 1964.

[13] G. Iooss and D. D. Joseph. Elementary Stability and Bifurcation Theory.

Springer, New York, 1980.

[14] M. Jakobson. Absolutely continuous invariant measures for a family of one-

dimensional maps. Comm. Math. Phys., 73:39–88, 1981.

[15] E. N. Lorenz. Deterministic non-periodic flow. J. Atmos. Sci., 20:130–141,

1963.

[16] J. Moser. Stable and Random Motions in Dynamical Systems. Ann. Math.
Studies 77. Princeton Univ. Press, Princeton NJ, 1973.

[17] L. Perko. Differential Equations and Dynamical Systems. Springer, New

York, 1996.
BIBLIOGRAPHY 147

About the author

Björn Birnir is a Professor of Mathematics at the University of California at
Santa Barbara (UCSB). He obtained his PhD in Mathematics at the Courant In-
stitute in 1981 and held positions at the University of Arizona Tucson, University
of California Berkeley and University of Iceland before coming to UCSB. He
served as the UCSB coordinator for nonlinear science from 1985-1990. He is
currently the director of the Center for Complex and Nonlinear Science at UCSB.
His current research interest are: Stochastic nonlinear partial differential equa-
tions and turbulence; dynamical systems theory of partial differential equations;
mathematical seismology and geomorphology; nonlinear phenomena in quantum
mechanical systems; complex and nonlinear models in biology and applications
of the above. The papers from the Center for Complex and Nonlinear Science can
be found at the website: http://repositories.cdlib.org/cnls/
Björn was the Chair of the UCSB Graduate Council 2004-2005 and he is cur-
rently (2008-2009) the Chair of the Council for Budget and Planning. He is an
active member of AMS, SIAM and AAAS and chaired the AMS-IMS-SIAM Se-
lection Committee for Summer Conference from 2004-2006. At UCSB he devel-
oped programs in Computational Science and in Applied Mathematics. Björn runs
a very active research program at UCSB where he has advised 25 Ph.D. students
and Postdoctoral Researchers and authored a large number of research papers and
monographs.